Model-Based Optimization of Combustion-Engine Control

Turesson, Gabriel

2018

Document Version:
Publisher's PDF, also known as Version of record

Link to publication

Citation for published version (APA):

Turesson, G. (2018). Model-Based Optimization of Combustion-Engine Control. [Doctoral Thesis (monograph),
Department of Automatic Control]. Department of Automatic Control, Lund Institute of Technology, Lund
University.

Total number of authors:

1

General rights
Unless other specific re-use rights are stated the following general rights apply:
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors
and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the
legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study
or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal

Read more about Creative commons licenses: https://creativecommons.org/licenses/

Take down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove
access to the work immediately and investigate your claim.

LUNDUNI
VERSI
TY

PO Box117
22100L und
+4646-2220000
Model-Based Optimization of
Combustion-Engine Control

Gabriel Turesson

Department of Automatic Control

The cover illustration shows heat-release data from a closed-loop experiment.

PhD Thesis TFRT-1120

ISBN 978-91-7753-705-2 (print)
ISBN 978-91-7753-706-9 (web)
ISSN 0280–5316

Department of Automatic Control

Lund University
Box 118
SE-221 00 LUND
Sweden

© 2018 by Gabriel Turesson. All rights reserved.

Printed in Sweden by Media-Tryck.
Lund 2018
Till Desire
Abstract

The work presented in this thesis is motivated by the need to reliably operate
a compression-ignition engine in a partially premixed combustion (PPC) mode.
Partially premixed combustion is a low temperature combustion concept, where
the ignition delay is prolonged to enhance fuel-air mixing in the combustion
chamber before the start of combustion. A premixed combustion process, in
combination with high levels of exhaust-gas recirculation (EGR), gives low com-
bustion temperatures, which decrease NOx and soot formation. Lowered com-
bustion temperatures also reduce heat-transfer losses which increase the ther-
modynamic engine efficiency. The ignition delay is, however, determined by che-
mical reactions rates, which are sensitive to temperature, gas-mixture composi-
tion, fuel properties and fuel-injection timing. This sensitivity makes PPC more
challenging to operate compared to conventional diesel combustion. Challenges
related to PPC include an increased sensitivity to operating conditions, decre-
ased combustion-timing controllability, high pressure-rise rates, and low com-
bustion efficiency at low engine loads. These challenges put high demands on
the engine control system that needs to be able to adjust fuel-injection timings
and durations to compensate for the combustion sensitivity.
Therefore, this thesis investigates closed-loop combustion control for relia-
ble PPC operation. The feedback loop from pressure-sensor measurement to
fuel-injection actuation is studied in particular. A common theme for the con-
trollers presented is the use of models in the controller design. Either to evaluate
controller performance in simulation, or to optimize engine performance online.
The principle of model predictive control is used for its ability to incorporate mo-
deled system behavior in the controller design, and to control multi-variable sys-
tems with input and output constraints.
The problem of tuning robust and noise insensitive combustion-timing con-
trollers, and its dependence on fuel reactivity is studied in simulation. Simulation
results reveal a nonlinear relation between start of injection and combustion ti-
ming that depends on both load and fuel reactivity. Optimization is used to find
robust and noise-insensitive controller gains. Guidelines for combustion-timing
controller tuning are also presented.

5
 Low-order autoignition models are evaluated and compared for the purpose
of model-based controller design. The comparison shows that a simple autoigni-
tion model is sufficient for control of the ignition delay when the cylinder-charge
properties are varied. This model is then used by a model predictive controller to
simultaneously control ignition delay and combustion timing in transient engine
operation, using both gas-exchange and fuel-injection actuation.
The effects of pilot injection on the combustion processes are characterized
experimentally. Experimental results show that a pilot injection can decrease the
main-injection ignition delay and maintain the pressure-rise rate at an accepta-
ble level. This is utilized by a presented fuel-injection controller that manages to
control both combustion timing and pressure-rise rate.
Strategies for improving the low-load performance of PPC are studied expe-
rimentally, where results show that the selection of injection timings and the use
of a pilot injection are important when maximizing the combustion efficiency.
The suggested low-load controller demonstrated a 9 % efficiency increase during
transient engine operation.
This thesis also investigates the design of a controller that utilizes the de-
grees of freedom enabled by multiple injections to efficiently fulfill constraints
on cylinder pressure, NOx emissions and exhaust temperature. A simulation
study shows a potential 2 - 4 % indicated efficiency increase when two injec-
tions are used instead of one. These findings motivated the design of a hybrid
multiple-injection controller that changes the number of injections depending
on operating conditions. The controller designed was capable of reproducing the
found efficiency increase experimentally with respect to constraints on pressure
and NOx emissions.
A model-predictive pressure controller is also introduced. The controller pre-
dicts how the cylinder pressure varies with fuel injection by taking advantage of
the estimated heat-release rate and a cylinder-pressure model. This feature was
used to adjust fuel-injection timings, durations, and number of injections, for
efficient constraint fulfillment in transient engine operation. Experimental re-
sults demonstrate that the pressure controller can also be used for tracking of
cycle-resolved in-cylinder pressure trajectories, as well as finding the most effi-
cient combustion timing.
Heat-release analysis is an essential component in the pressure-sensor feed-
back loop. Methods for calibrating heat-release model parameters with the use
of engine data, and methods for detecting combustion timings are therefore dis-
cussed in the thesis.
The experimental results presented were conducted on a heavy-duty Scania
D13 engine with a modified gas-exchange system. The fuel used was a mixture
(by volume) of 80 % gasoline and 20 % n-heptane, to elevate the fuel octane num-
ber and allow for longer ignition delays.

6
Acknowledgments

I would like to thank my supervisor, Rolf Johansson, for his consistent support
and advice, for sharing his knowledge of scientific work, and for setting high stan-
dards in the project. My co-supervisor Per Tunestål has been very helpful with
discussions and suggestions on theoretical and practical matters related both to
combustion engines and control. Bo Bernhardsson gave interesting courses in
control and optimization that contributed to the project.
The experimental work was done in collaboration with my colleague, Lian-
hao Yin. His knowledge about, and experience with laboratory work and pro-
gramming made it possible for us to quickly build the test rig and conduct the
experiments presented in the thesis. The technicians at the Division of Combus-
tion Engines were very helpful in keeping the engine running and re-building the
experimental platform when necessary. I am also grateful to Ola Stenlåås for ta-
king a great interest in our project and for giving constructive feedback on our
progress. I wish to express my gratitude to the colleagues who took their time
and helped me with the proofreading of this thesis.
I also extend my acknowledgments to colleagues at the Department of Auto-
matic Control and the Department of Energy Sciences for interesting discussions
on research and related subjects, and to the technical and administrative staff at
the departments for assistance on various subjects related to my research. Spe-
cial thanks goes to Leif Andersson for his support in the finalization of this thesis.
Finally, I would like to thank my wife Desire, my family, and my friends for
their encouragement and support.

Financial Support
The author would like to acknowledge the Competence Center for Combus-
tion Processes (KCFP) and the Swedish Energy Agency for the financial support
(project number 22485-3), Scania for supplying the experimental engine, and the
KFCP PPC Control reference group for consistent feedback on the work. The au-
thor is also a member of the LCCC Linnaeus Center and the ELLIIT Excellence
Center at Lund University.

7
Contents

Nomenclature 13
1. Introduction 17
1.1 Combustion-Engine Challenges . . . . . . . . . . . . . . . . . . . 17
1.2 Fundamental Engine Principles . . . . . . . . . . . . . . . . . . . 18
1.3 Low Temperature Combustion . . . . . . . . . . . . . . . . . . . . 23
1.4 Engine Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.5 Outline and Contributions . . . . . . . . . . . . . . . . . . . . . . 31
2. Modeling 37
2.1 Control-Oriented Modeling . . . . . . . . . . . . . . . . . . . . . . 37
2.2 In-Cylinder Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.3 Gas-Exchange System . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.4 Ignition-Delay Modeling . . . . . . . . . . . . . . . . . . . . . . . 50
2.5 NOx -formation Model . . . . . . . . . . . . . . . . . . . . . . . . . 66
3. Control and Estimation Methods 70
3.1 Control Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.2 State Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4. Heat-Release Analysis 77
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.3 Filter Configurations . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5. Experimental Setup 96
5.1 The Scania D13 Engine . . . . . . . . . . . . . . . . . . . . . . . . 96
5.2 Instrumentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.3 Control-System Architecture . . . . . . . . . . . . . . . . . . . . . 100

9
Contents

6. Proportional-Integral Combustion-Timing Control 102

6.1 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.2 Steady-State Characteristics . . . . . . . . . . . . . . . . . . . . . 104
6.3 Controller Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7. Model-Based Control of Combustion Timing and Ignition Delay 123
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.2 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.3 Model Predictive Control Formulation . . . . . . . . . . . . . . . 129
7.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.5 Discussion and Conclusions . . . . . . . . . . . . . . . . . . . . . 134
8. Pilot Injection 140
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
8.2 Experimental Characterization . . . . . . . . . . . . . . . . . . . . 144
8.3 Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.4 Controller Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 153
8.5 Discussion and Conclusions . . . . . . . . . . . . . . . . . . . . . 154
9. Low-Load Control 158
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
9.2 Low-Load Experiments . . . . . . . . . . . . . . . . . . . . . . . . 159
9.3 Suggested Controller . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.4 Controller Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 168
9.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
10. Constraint Handling with Multiple Injections 173
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
10.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . 176
10.3 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
10.4 Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
10.5 Experimental Evaluation . . . . . . . . . . . . . . . . . . . . . . . 190
10.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
10.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
11. Pressure Prediction and Efficiency Optimization 202
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
11.2 Controller Description . . . . . . . . . . . . . . . . . . . . . . . . 203
11.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
11.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
11.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
12. Predictive Constraint Handling and Pressure Tracking 220
12.1 Model Predictive Control Formulation . . . . . . . . . . . . . . . 221
12.2 Modeling and Heat-Release Detection . . . . . . . . . . . . . . . 226

10
 Contents

12.3 Adding and Removing Injections . . . . . . . . . . . . . . . . . . 233

12.4 Constraint Handling . . . . . . . . . . . . . . . . . . . . . . . . . . 236
12.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . 236
12.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
12.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
13. Conclusions and Future Research 250
Bibliography 253

11
Nomenclature

Symbol Descriptions
The table below summarizes the most frequently used notation in the thesis.

Notation Description

A Cylinder area
B Cylinder bore
C1 & C2 Heat-transfer coefficients
cp Specific heat at constant pressure
cv Specific heat at constant volume
γ Ratio of specific heats
d p max Maximum cylinder pressure derivative
Ea Activation energy
η GIE Gross indicated efficiency
η NIE Net indicated efficiency
η th Thermodynamic efficiency
λ Relative air/fuel ratio
φ Fuel/air equivalence ratio
ṁ air Air mass flow
ṁ EGR EGR mass flow
ṁ f Fuel mass flow
mf Fuel mass
NOx Oxides of nitrogen
Nspeed Engine speed
p In-cylinder pressure
p ex Exhaust-manifold pressure
p in Intake-manifold pressure
p IMEPg Gross indicated mean effective pressure
p IMEPn Net indicated mean effective pressure

13
Nomenclature

p PMEP Indicated pumping losses

p max Maximum cylinder pressure
p rail Common-rail pressure
dQ c /d θ Heat-release rate
dQ ht /d θ Heat-transfer rate
Q LHV Lower heating value
R2 Coefficient of determination
R̃ Universal gas constant
R Specific gas constant
rc Compression ratio
r EGR EGR ratio
rp Pilot ratio
Sp Mean piston speed
T In-cylinder temperature
Tc Coolant temperature
TEVO Temperature at exhaust-valve opening
TIVC Temperature at intake-valve closing
Tex Exhaust-manifold temperature
Tin Intake-manifold temperature
Tres Residual-gas temperature
Tw Wall-surface temperature
θ Crank angle
θCT Combustion timing
θDOI Fuel-injection duration
θEOC End of combustion
θSOC Start of combustion
θSOI Fuel-injection timing
∆θTDC TDC offset
θx Crank-angle of x % burnt
θHP High-pressure EGR valve
θLP Low-pressure EGR valve
θcool Cool-path FTM valve
θhot Hot-path FTM valve
τ Ignition delay
u System input
Vd Displacement volume
V Cylinder volume
x System state
xc Constraint for x
xr Set point for x
[x] Concentration of species x
y System output
ω Mean gas velocity

14
 Nomenclature

Abbreviation Description

0D Zero dimensional
BC Boundary condition
CAD Crank angle degree
CDC Conventional diesel combustion
CFD Computational fluid dynamics
DOI Duration of injection
EGR Exhaust gas recirculation
EVO Exhaust valve opening
EKF Extended Kalman filter
FTM Fast thermal management
HC Hydrocarbon
HCCI Homogeneous charge compression ignition
IAE Integrated absolute error
ICE Internal combustion engine
IMEP Indicated mean effective pressure
IVC Inlet valve closing
KF Kalman filter
LTC Low temperature combustion
MPC Model predictive control
MVM Mean value model
NTC Negative temperature coefficient
ON Octane number
PF Particle filter
PI Proportional integral
PPC Partially premixed combustion
PM Particulate matter
PRF Primary reference fuel
rpm Revolutions per minute
RMSE Root mean square error
SOC Start of combustion
SOI Start of injection
TDC Top-dead-center
QP Quadratic program
VI Virtual instrument

15
1
Introduction

Combustion-engine technology has made tremendous advances during the last

decades due to increasingly stringent demands for a reduction in both emis-
sion levels and fuel consumption. These demands have necessitated an incre-
ase in the number of engine sensors and actuators [Isermann, 2014]. At the
same time, there has been an exponential growth in electronics, and the avai-
lability of computational power has increased [Leen and Heffernan, 2002; Broy
et al., 2007]. These trends have opened up new possibilities for more intelligent
engine-control systems, capable of monitoring the combustion process, and in
real time, optimize actuator adjustments for clean and efficient combustion.
Such control systems could be used to enable sensitive low-emission combus-
tion concepts. Partially premixed combustion is a combustion concept that has
previously been shown to provide high efficiency and low emissions simultane-
ously. Its sensitivity to chemical reaction rates has, however, made it challenging
to operate. This thesis presents work on how to operate a compression-ignition
engine in a partially-premixed combustion mode, with model-based control and
pressure-sensor feedback. This thesis also investigates how feedback control can
be used to optimize fuel injection for efficient fulfillment of constraints on cy-
linder pressure, NOx emissions and exhaust temperature. Such methods could
also be used to improve performance of conventional combustion concepts. The
controllers and methods presented were evaluated experimentally using a modi-
fied heavy-duty engine, which was run on a high octane-number fuel to prolong
ignition delays and enhance premixing of fuel and air.
This chapter provides background and motivation for the work presented in
this thesis together with a description of the research contributions made.

1.1 Combustion-Engine Challenges

The internal combustion engine has revolutionized transportation since the
19th century. Today, the world’s vehicle population consists of more than one
billion cars, with a yearly production of 95 million, of which the vast majority

17
Chapter 1. Introduction

is propulsed by a combustion engine [Sperling and Gordon, 2009; OICA, 2017].

The reason for its persistent popularity is a combination of simplicity, durabi-
lity, low power-to-weight ratio, high power controllability and reasonable effi-
ciency [Heywood, 1988]. Engine research and development are mainly focused
on legislation-enforced emission reduction and the aim to reduce fuel consump-
tion. This is due to toxic local emissions [Haagen-Smit, 1952; Sher, 1998], and the
fact that the transportation sector contributes to 23 % of global green-house gas
emissions [IPCC, 2014]. The European NOx and soot-particle emission limits for
heavy-duty engines during the past 20 years are presented in Fig. 1.1. Emissions
of CO, unburned hydrocarbons (HC) and particle number count are also regula-
ted. Forthcoming legislation is expected to include standards for greenhouse-gas
emission levels as the European Union targets a 60 % greenhouse-gas emission
reduction in the transportation sector by 2050, compared to the levels of 1990
[EC, 2016]. Battery-electric and fuel-cell vehicles are possible zero-emission
alternatives to the internal combustion engine [Contestabile et al., 2011].
However, considering the large number of combustion engines on our roads
today, mass production will probably continue for several years, especially in the
heavy-duty sector [McKinsey, 2014; Askin et al., 2015]. In this scenario, techno-
logical combustion-engine advances in combination with hybridization and the
usage of biofuels could constitute a cost-effective path towards reduced local
and global emissions [Enkvist et al., 2007; Johansson et al., 2013].
Engine-technology advances for cleaner combustion includes high-pressure
fuel-injection, exhaust-gas recirculation (EGR), and compression and cooling of
inducted air [Majewski et al., 2006]. Exhaust filters and catalysts are also neces-
sary for fulfillment of current legislation. The resulting engine complexity has led
to an increase in the number of sensors and actuators which has promoted the
development of more advanced engine-control systems. Focus has also been
directed towards the development of new, clean and efficient combustion con-
cepts. A number of such concepts utilize low temperature combustion, with low
emission formation and heat-transfer rates, achieved through increased mixing
of fuel and air in the combustion chamber. The following sections give a brief
overview of fundamental engine principles and compare conventional diesel
combustion with low temperature combustion.

1.2 Fundamental Engine Principles

The reciprocating internal combustion engine is a heat engine where combus-
tion of fuel and air occurs inside a cylinder. The combustion fluids perform ex-
pansion work on a piston whose linear movement is converted to rotation of a
crankshaft. The basic geometry of an engine cylinder is shown in Figure 1.2. The
work presented in this thesis concerns the four stroke compression-ignition en-
gine. The four-stroke cycle starts with air induction through the intake valves

18
 1.2 Fundamental Engine Principles

Euro I-VI NOx and PM Legislation Euro I

Euro II
Euro III
0.6 Euro IV
Euro V
Euro VI
PM [g/kWh]

0.4

0.2

0
0 2 4 6 8
NOx [g/kWh]

Figure 1.1 Legislated emission levels for heavy-duty vehicles in the European
Union (Euro I-VI) during the past 20 years [EU, 2007]. The emission goals have
been met by gradual improvements of emission control techniques and improved
fuel quality. Improvements from Euro IV were done with the help of exhaust-gas
after treatment.

due to downward motion of the piston. The air is then compressed during the
compression stoke and fuel is injected as the piston approaches top-dead-center
(TDC). At this point, the temperature is sufficiently high for autoignition to oc-
cur. Combustion leads to a pressure increase that generates work during the ex-
pansion stroke. Combustion products are then scavenged as the piston moves
upward with the exhaust-valves open during the exhaust stroke. The four stro-
kes and corresponding cylinder pressure and volume curves are presented in
Figs. 1.3 and 1.4. The cylinder pressure was obtained from the engine used in
the experimental work presented in the thesis, whereas the volume was compu-
ted from the cylinder geometry. Fuel injection is indicated by the injector-current
pulse located before TDC.

Conventional Diesel Combustion

The following description of conventional diesel combustion (CDC) is based on
the conceptual model presented by Dec [1997].
Conventional diesel combustion is initiated by high-pressure fuel injection
into a compressed, hot (∼1000 K) air charge, close to TDC. The injected fuel pro-
pagates into the combustion chamber and forms a conical jet of fuel droplets.
The fuel jet becomes increasingly diluted with hot air and vaporizes as it ex-
pands. After a certain traveling distance along the jet axis (∼20-25 mm), called

19
Chapter 1. Introduction

injector
intake valve exhaust valve

TDC
combustion
chamber
piston

BDC

connecting rod θ

crankshaft

Figure 1.2 The basic geometry of an engine cylinder. Combustion of fuel and
air in the combustion chamber leads to a pressure increase that generates linear
piston motion. The linear motion is converted to rotational motion of the crank-
shaft. Flow of fuel and air are governed by poppet valves, a fuel injector and the
motion of the piston. The acronyms TDC and BDC stand for top-dead-center
and bottom-dead-center, indicating the top and bottom positions of the piston.
Crank-angle degree is denoted θ.

the lift-off length, the fuel has vaporized completely. Chemical reactions are ini-
tiated all over the jet cross section after further air entrainment. Initial reactions
are followed by rapid, rich, premixed combustion and the resulting temperature
increase leads to formation of soot due to an excess of fuel. Air entrainment con-
tinues as the reacting fuel travels along the spray axis and a quasi-steady diffu-
sion flame is formed along the jet periphery when stoichiometric conditions are
reached. At this stage, the combustion rate is controlled by how fast the injected
fuel is vaporized, mixed with air and supplied to the diffusion flame. This type
of combustion is therefore referred to as mixing controlled. Temperature reaches
its maximum in the vicinity of the flame which causes nitrogen to oxidize and
form harmful NOx emissions. High temperature in combination with availability
of oxygen also result in soot oxidation which gives the characteristic diesel-flame
luminosity. Furthermore, these conditions lead to almost complete oxidation of

20
 1.2 Fundamental Engine Principles

Intake Compression Expansion Exhaust

Figure 1.3 The principle of a four-stroke compression-ignition engine. Air is first

inducted from the intake manifold during the intake stroke. The air is then com-
pressed during the compression stoke and fuel is injected as the piston approa-
ches TDC, when the cylinder temperature is sufficiently high for autoignition to
occur. The fuel autoignites shortly after TDC and the resulting pressure increase
generates work during the expansion stroke. The combustion products are then
scavenged during the exhaust stroke, as the piston moves upward with the ex-
haust valves open.

CO and HC during the expansion stroke. This conceptual description of CDC is

illustrated in Fig. 1.5, where two jet intersections are presented.
The left diagram in Fig. 1.5 shows a fuel jet 5 crank-angle degrees (CAD) af-
ter injection, where the liquid fuel has started to vaporize. Initial rich premixed
combustion is indicated in purple. The diagram to the right shows the same jet
1.5 CAD later. A hot diffusion-flame front (green) has been established and sur-
rounds a region of soot formation due to rich and hot conditions.
The temperature T , and the ratio of the actual fuel/air ratio to the stoichio-
metric fuel/air ratio φ have significant effect on soot formation and oxidation.
Temperature also affects formation of NOx . Emission characteristics can there-
fore be studied in a φ − T diagram, where emission-formation level curves are
presented as a function of φ and T . Such diagrams have been generated from ex-
perimental data, and by simulating soot and NOx concentrations at fixed φ and T
for a given residence time [Aoyagi et al., 1980; Kamimoto and Bae, 1988; Kitamura
et al., 2002; Kook et al., 2005]. A φ − T diagram for combustion of n-heptane, a
fuel with similar properties to diesel, is presented in Fig. 1.6.

21
Chapter 1. Introduction

Cylinder Pressure
100
p [bar] Intake Compression Expansion Exhaust

fuel injection
50

0
−300 −200 −100 0 100 200 300
θ [CAD]

Cylinder Volume
·10−3

2
V [m3 ]

0
−300 −200 −100 0 100 200 300
θ [CAD]

Figure 1.4 Cylinder pressure and volume during the engine cycle. The cylinder
pressure was measured from the engine used for the experimental work presented
in this thesis, whereas the volume was computed from the cylinder geometry. Fuel
injection is indicated by the injector current pulse located before TDC.

The journey of a fuel particle along the diesel-jet axis is indicated by the
orange trajectory in this diagram. The fuel/air mixture starts rich (high φ) with
fairly low temperature after vaporization (1). Premixed rich combustion is then
initiated as the fuel mixes with air (φ ∼ 2 − 4). Temperature increases steeply and
the rich, high-T , soot-formation region is reached (2). The fuel becomes incre-
asingly diluted and the temperature peaks at the diffusion flame boundary at
close to stoichiometric conditions (3). The lean, high-T region is favorable for
thermal NOx formation and the availability of oxygen promotes soot oxidation
as the burned mixture cools down during the expansion stroke (4). The inter-
section of the NOx -formation zone with the soot-oxidation zone leads to a fun-
damental diesel-combustion trade-off between NOx and soot emissions. If NOx
formation is to be avoided, soot oxidation will simultaneously decrease with hig-
her soot-emission levels as a result.
It should be noted, however, that this conceptual model only gives a quali-
tative understanding of diesel-combustion characteristics. In reality, there is a
distribution of φ and T around the trajectory in Fig. 1.6, [Kitamura et al., 2002].

22
 1.3 Low Temperature Combustion

Distance from injector [mm]

0 10 20 30 40 50 0 10 20 30 40 50

5◦ after injection 6.5◦ after injection

Liquid Fuel First-Stage Ignition Second-Stage Ignition of

(H2 CO, H2 O2 , CO, UHC) Fuel-Rich Mixtures
Pre-Ignition Vapor Fuel Second-Stage Ignition of Soot or Soot Precursors (PAH)
Intermediate Stoichiometry
or Diffusion Flame (OH)

Figure 1.5 Conceptual model of conventional diesel combustion. The left dia-
gram shows a fuel jet 5 crank-angle degrees (CAD) after injection, where liquid
fuel has vaporized and mixed with air. Initial premixed combustion is indicated
in purple. The diagram to the right shows the same jet 1.5 CAD later, where a hot
diffusion-flame front (green) has been established. Source: [Musculus, 2006].

1.3 Low Temperature Combustion

Combustion concepts with improved paths through the φ−T diagram have been
under intense study during the past decades. Low temperature combustion (LTC)
concepts utilize enhanced fuel-air mixing to increase the combustion-zone heat
capacity. The increased heat capacity lowers combustion temperatures and ma-
kes it possible to avoid the emission-formation regions in Fig. 1.6.

HCCI
The interest in LTC emerged with the discovery of homogeneous charge com-
pression ignition (HCCI). A concept where a diluted homogeneous charge is in-
ducted and autoignitioned by compression. It was first studied in two-stroke en-
gines [Onishi et al., 1979] and was later shown to yield low emission levels in com-
bination with high efficiencies in the low-load operating region of a four-stroke
engine [Epping et al., 2002]. The blue line in Fig. 1.6 indicates the path taken
by a fuel element in the case of ideal HCCI combustion. It is assumed that the
fuel is completely mixed with air prior to the start of combustion, φ < 1. The
HCCI trajectory shows lean, premixed combustion that avoids the soot forma-

23
Chapter 1. Introduction

5
25 % Soot Formation

4
Local Equivalence Ratio, φ [-]

(1)
3
1%

(2)
Diesel Combustion
2

500 ppm
PPC
(3)
1
HCCI (4)
5000 ppm
NOx Formation
Soot Oxidation
0
600 1,000 1,400 1,800 2,200 2,600 3,000
Local Temperature, T [K]

Figure 1.6 Emission formation as a function of φ and T . Combustion paths

for HCCI, PPC and conventional diesel combustion are indicated by the blue,
green and orange lines. The increased dilution of fuel with air in HCCI and PPC
gives a higher combustion-zone heat capacity. This reduces T and φ, and the
emission-formation zones can be avoided. Source: [Kitamura et al., 2002; Kook
et al., 2005].

tion region. Furthermore, the reduced temperature results in low NOx -emission
levels. In HCCI, the combustion timing is completely determined by chemical
autoignition kinetics as compared to injection-controlled diesel combustion.
Combustion timing can therefore only be controlled by varying the tempera-
ture and mixture composition during the compression stroke. This poses a chal-
lenge with respect to combustion-timing sensitivity and controllability. Alterna-
tive combustion-timing strategies, such as variable-valve timing [Agrell et al.,
2003], variable compression ratio [Haraldsson et al., 2002] and dual-fuel opera-
tion [Olsson et al., 2001] have been proposed to control the combustion timing in
HCCI. Another challenge with HCCI is the limited operating range due to violent
combustion rates during high-load operation [Olsson et al., 2004].

24
 1.3 Low Temperature Combustion

PPC
Low temperature combustion can also be obtained by prolonging the ignition
delay of conventional diesel combustion. This can be done by diluting the in-
ducted air charge with EGR, and by injecting fuel earlier during the compression
stroke, or later during the expansion stroke when the temperature is lower [Ta-
keda and Keiichi, 1995; Kimura et al., 1999; Akihama et al., 2001]. With direct in-
jection, the fuel-air mixture obtained is not as homogeneous as in HCCI. On the
other hand, combustion timing is only to a limited degree controlled by chemical
kinetics. Emission levels can therefore be reduced with maintained combustion
controllability. These concepts are commonly referred to as partially premixed,
and has been applied by means of different techniques, leading to a number of
names and abbreviations existing in the literature. Examples are MK, PCCI and
RCCI [Kimura et al., 1999; Kanda et al., 2005; Kokjohn et al., 2011]. This thesis ad-
dresses topics related to a concept called partially premixed combustion (PPC).
The conceptual model developed by Musculus [2006] is summarized below for
the purpose of describing PPC.
In PPC, the fuel takes an intermediate path through the φ − T diagram. This
is illustrated by the green line in Fig. 1.6. Injection occurs earlier during the com-
pression stroke, which increases the fuel-traveling distance prior to vaporization
due to reduced mixture temperature and density. The reduced temperature and
increased EGR dilution extend the ignition delay and make it possible for the
majority of the fuel to vaporize before the start of combustion. Increased mix-
ing leads to a more spatially distributed fuel jet as compared to CDC. Combus-
tion is therefore initiated more uniformly over the jet cross section. The incre-
ased combustion-zone heat capacity reduces combustion temperature, and the
simultaneous reduction of φ and T reduces formation of both NOx and soot. Two
PPC fuel-jet intersections are presented in Fig. 1.7. Here, the fuel starts to react
later (12 CAD after injection) as compared to Fig. 1.5, which allows for complete
fuel vaporization before the start of combustion. The fuel jet also occupies a lar-
ger region during combustion. This leads to an increased dilution with smaller
regions of soot formation.
It has been proven difficult to obtain sufficient ignition delay for premixed
combustion with diesel fuels at high-load conditions [Noehre et al. 2006]. This
could be remedied by increasing the fuel autoignition resistance. Gasoline PPC
was introduced by Hildingsson et al. [2006] who showed that longer ignition de-
lays could be achieved even at high-load conditions. At the author’s engine lab-
oratory, Manente et al. [2010c] showed that gasoline PPC could achieve gross in-
dicated efficiencies between 52 and 55 % from idle to 26 bar (indicated mean
effective pressure), with Euro VI compatible emission levels. This was achieved
with EGR ratios at 50 %, φ at 0.75, and an advanced fuel-injection strategy. Simu-
lation results in [Fridriksson et al., 2011] attributed the high efficiency obtained
to low heat-transfer losses resulting from low combustion temperatures. These

25
Chapter 1. Introduction

Distance from injector [mm]

0 10 20 30 40 50 0 10 20 30 40 50

12◦ after injection 14◦ after injection

First-Stage Ignition Intermediate Ignition Second-Stage Ignition of
(H2 CO, H2 O2 , CO, UHC) (CO, UHC) Fuel-Rich Mixtures
Pre-Ignition Vapor Fuel Second-Stage Ignition of Soot or Soot Precursors (PAH)
Intermediate Stoichiometry
or Diffusion Flame (OH)

Figure 1.7 Conceptual model of low load, single-injection, and EGR-diluted par-
tially premixed combustion (PPC). In PPC, the fuel reacts later compared to con-
ventional diesel combustion, see Fig. 1.5. This allows for complete fuel vapori-
zation before the start of combustion. The fuel jet also occupies a larger region
during combustion, which leads to increased dilution with smaller regions of soot
formation. Source: [Musculus, 2006].

results were the main motivation for the work presented in this thesis.
The increased dependency on chemical kinetics makes PPC more sen-
sitive to operating conditions, as compared to CDC. Variations in tempera-
ture, dilution and fuel reactivity might lead to undesired combustion in the
emission-formation regions in Fig. 1.6. Other PPC challenges include violent
combustion rates, increased cylinder-to-cylinder variation, and misfire if the
fuel does not ignite properly. This thesis therefore studies the problem of con-
trolling PPC. With the objective of advancing the concept from manual operation
in steady state, towards autonomous and transient operation in a multi-cylinder
engine.

1.4 Engine Control

Automatic control concerns automatic operation and regulation of systems. For
most control problems, the state of the system to be controlled x is influenced
by a control input u, and information about x can be obtained from a measured
output y. Controller design addresses the problem of deciding u in order for the

26
 1.4 Engine Control

Disturbance

r P e u y
Controller Engine

−1

Figure 1.8 A system controlled in closed loop. The controller decides u based on
deviation between the measured system output y and a set point r .

specified system-performance requirements to be fulfilled. Engine-performance

requirements are related to efficient, durable and reliable operation subject to
constraints on emission and noise levels. At a price that is competitive for mass
production.
Compression-ignition engines have traditionally been controlled in open-loop,
where the injected fuel amount has been determined by the accelerator-pedal
position, engine speed and air-fuel ratio as opposed to generated work output.
Fuel-injection timings are commonly computed from calibrated maps [Guzzella
and Onder, 2009]. In closed-loop control, the control action is dependent on the
measured process output, see Fig. 1.8. Closed-loop controllers have the advan-
tage of making the system more resilient to external disturbances and variation
in system components. Examples related to combustion engines are variations
in fuel and gas-mixture properties, driving patterns, and aging of hardware com-
ponents [Saracino et al., 2015]. Potential disadvantages with closed-loop control
are possible dynamic instabilities and the introduction of sensor noise into the
system. This imposes controller-design challenges in terms of trade-offs between
achievable system performance and robustness [Åström and Murray, 2010].
The work presented in this thesis investigates how timings and durations of
fuel-injection pulses should be decided to control the combustion processes with
the use of in-cylinder pressure measurement. For this control problem, the main
actuator is a solenoid injector, connected to a common-rail fuel system. The in-
jected fuel quantity is determined by the common-rail pressure and the opening
duration of the injector nozzle. Injector-opening timings and durations are de-
termined by current pulses sent to the injector solenoid. The possibility to divide
the fuel among several injection events gives additional degrees of freedom for
combustion control. A detailed description of the workings of a solenoid injec-
tion is provided in [Bosch, 2011].
The cylinder pressure is perhaps the most fundamental variable available for
direct measurement in an internal combustion engine. In this thesis it is measu-

27
Chapter 1. Introduction

80 p

p [bar] 60

40 injector current

20 dQ c /d θ

0
−30 −20 −10 0 10 20 30 40
θ [CAD]

Figure 1.9 This thesis investigates the problem of how to decide timings and
durations of fuel-injection pulses to control the combustion processes and the
in-cylinder pressure p. This is a control problem with several degrees of freedom
and a highly informative measured system output as feedback signal.

red with piezoelectric pressure transducers, mounted in the cylinder head. These
sensors utilize the piezoelectric effect where a charge is generated when a pie-
zoelectric crystal is exposed to a force. The measured in-cylinder pressure can
be used to compute indicated engine work, heat-release rate and NOx forma-
tion. The heat-release rate dQ c /d θ can be used to compute combustion timings
and ignition delays, which are important indicators for efficiency and emission
formation. Although cylinder pressure sensing is a widely used in engine rese-
arch, development, and calibration, in-cylinder pressure sensors have not yet
reached widespread use in production vehicles due to high technical demands
and associated cost. Recent announcements still indicate that cylinder pressure
sensing might be used in future production vehicles [BorgWarner, 2014; Nagatsu
et al., 2017]. There are several reviews describing the potential use for pressure
sensors in engine control and diagnostics, see [Powell, 1993; Iorio et al., 2003;
Eriksson and Thomasson, 2017].
Cycle-resolved input and output signals are presented as a function of θ in
Fig. 1.9. The figure shows injector-current pulses, the measured in-cylinder pres-
sure p and the computed heat-release rate dQ c /d θ.
Favorable cylinder boundary conditions in terms of intake temperature, pres-
sure and composition were in this work obtained by regulating mass flows
through EGR paths and a charge-air cooler. This was done by actuating valve po-
sitions in the gas-exchange system. Sensors in the gas-exchange system measu-
red temperatures, pressures, air-mass flow and exhaust oxygen concentration. A
more detailed description of the experimental setup and the gas-exchange sys-
tem is given in chapters 2 and 5.

28
 1.4 Engine Control

Control of Partially Premixed Combustion

Partially premixed combustion is sensitive to the cylinder-mixture temperature
and composition. This puts high demands on the engine-control system, where
actuators have to be accurately set for satisfactory performance. A feedback
loop from measured cylinder pressure to fuel injection could significantly reduce
combustion sensitivity. The foremost objective of this feedback loop is to reduce
combustion-timing sensitivity through injection-timing adjustment for efficient
delivery of the desired work output. If combustion occurs too early during the
compression stroke, pressure buildup would contribute negatively to the produ-
ced work output. If combustion on the other hand occurs too late, there is a risk
of incomplete combustion as a result of too lean and cold fuel-air mixtures. It is
also of interest to regulate the ignition delay to ensure sufficient fuel-air mixing.
Too high combustion rates is another known issue with PPC and other pre-
mixed combustion concepts. This problem can be solved by introducing a pi-
lot injection that reduces the ignition delay and the main-injection combustion
rate. Feedback control can be used to ensure a sufficient pilot amount for accep-
table combustion rates. These PPC-related control challenges are illustrated in
Fig. 1.10 together with performance improvements achieved with fuel-injection
adjustment. These challenges and corresponding feedback-controller designs
will be described in greater detail in chapters 6 to 9.

Optimal Control
A common theme for the work presented in this thesis is to represent control
problems as mathematical optimization problems with the system input u as the
optimization variable

minimize J (u, x) (1.1)

u
subject to g (u, x, y) = 0
f (u, x) ≤ 0

The cost function J is used to represent controller objectives such as set-point

tracking and minimization of fuel consumption. The equality constraint deter-
mines the relation between input, state and output and is a model of the sys-
tem to be controlled. The function g can also incorporate system dynamics by
including time derivatives or difference equations with respect to x. The inequa-
lity constraint describes constraints and limitations with respect to x and u. This
problem representation is convenient for engine-control purposes where it is of-
ten desirable to minimize fuel consumption and load set-point deviation without
violating constraints with respect to cylinder pressure, emissions and actuator li-
mitations. High engine efficiency typically demands operation close to, or on the
boundary of admissible x and u.

29
Chapter 1. Introduction

100 100
open loop closed loop
a) b)
80 80
p [bar]

p [bar]
60 60

40 40

20 20
−10 0 10 20 −10 0 10 20
θ [CAD] θ [CAD]

100 100
misfire
c) d)
80 80
p [bar]

p [bar]
60 60

40 40
knock
20 20
−10 0 10 20 −10 0 10 20
θ [CAD] θ [CAD]

Figure 1.10 This thesis investigates the use of closed-loop control for
reliable PPC operation. Presented controller designs manage to: Reduce
cylinder-to-cylinder variation for consistent and efficient work output with suf-
ficient ignition delay; Compare (a) where same injection durations and timing
are actuated to the different cylinder, with (b), where work output and com-
bustion timings are regulated in closed loop; Control the pressure-rise rate with
pilot-injection adjustment to avoid knock and maintain an efficient combustion
timing (c); Improve low-load operation by adjusting intake conditions and fuel
injection to avoid misfire and incomplete combustion (d).

The approach taken here was to let the controller repeatedly solve (1.1) on
a cycle-to-cycle basis with respect to measured y and a time horizon of future
inputs, states and outputs. This optimal-control technique is called model pre-
dictive control (MPC) and has gained attention in several context, for example
process control, automotive applications and combustion-engine control. Mo-
del predictive control will be described in greater detail in Chapter 3.
Chapters 10 to 12 of this thesis investigates the optimization problem of how
pressure-sensor feedback and actuation of a number of fuel injections could
be combined to efficiently fulfill constraints on cylinder pressure, NOx forma-
tion and exhaust-gas temperature. The aim of this investigation was to design a
controller capable of automatically finding efficient fuel-injection timings, du-
rations and number of injections, as a function of the engine operating point.

30
 1.5 Outline and Contributions

Predictive Pressure Control

Measurement
100 Prediction
p
p limit Constraint

80
NOx limit NOx
p [bar]

60

40

20
dQ c /d θ

0
−20 −10 0 10 20 30 40
θ [CAD]

Figure 1.11 Fuel-injection optimization for efficient fulfillment of pressure and

NOx -emission constraints is studied in this thesis. This figure illustrates the pre-
dictive controller technique introduced in chapters 11 and 12. The controller pre-
dicts how future-cycle pressure (dashed) varies with fuel injection with the use
of a pressure model and the previous-cycle pressure measurement (solid). The
prediction is used to optimize fuel-injection timings and durations so that con-
straints on p and NOx are efficiently fulfilled (dash-dotted).

Two types of controllers were studied. The first controller was designed with sim-
ple proportional/integral controller components and heuristic constraint hand-
ling. The second controller employed the MPC principle and utilized methods for
cylinder-pressure approximation and heat-release analysis to predict how the cy-
linder pressure varies with fuel-injection parameters. This principle is illustrated
in Fig. 1.11, where a controller predicts future-cycle fuel-injection adjustment
and cylinder-pressure variation (dashed) based on a pressure model and the me-
asured pressure from the previous cycle (solid). This allows the controller to re-
peatedly optimize fuel-injection timings and durations on a cycle-to-cycle ba-
sis, so that the predicted engine outputs efficiently fulfill specified constraints
(dash-dotted).

1.5 Outline and Contributions

The thesis begins with two chapters that introduce the models and control prin-
ciples used. The introductory chapters are followed by chapters on heat-release

31
Chapter 1. Introduction

analysis and the experimental platform. The main contributions of the thesis are
presented in chapters 6 to 12, where chapters 6 to 9 cover engine experiments
and controller designs related to partially premixed combustion. Chapters 10 to
12 are focused on fuel-injection optimization and constraint fulfillment, where
the results are also applicable to conventional compression-ignition engines.
A more detailed description of the chapters are given below along with refe-
rences to publications on which they are based. Preliminary versions of parts of
the research presented in this thesis was published in the licentiate thesis by the
author:

Ingesson, G. (2015). Model-Based Control of Gasoline Partially Premixed Combus-

tion. Licentiate Thesis TFRT-3268. Dept. of Automatic Control, Lund Univer-
sity, Lund, Sweden.

Note that the author’s previous surname was Ingesson.

Chapter 2
This chapter presents models used for simulation, state estimation and control-
ler design. It includes control-oriented models of the gas-exchange system and
in-cylinder processes. Model-calibration results are also presented. The main
contribution of this chapter is an evaluation of low-order ignition-delay models
for the purpose of control. The results show that a fairly simple model can be
used to predict the relation between intake conditions and ignition delay. The re-
lation between injection timing and ignition delay when the gain from injection
timing to ignition delay changes sign was, however, not adequately captured by
the models considered.

Related Publication
Ingesson, G., L. Yin, R. Johansson, and P. Tunestål (2017). “An investigation on
ignition-delay modeling for control”. Int. J. Powertrains 6:3, pp. 282–306.

Chapter 3
The control principles used are presented in this chapter with emphasis on mo-
del predictive control. This chapter also presents the state-estimation methods
used to estimate heat-release model parameters and EGR mass flow in subse-
quent chapters.

Chapter 4
This chapter introduces heat-release analysis methods used to extract combus-
tion information from in-cylinder pressure measurement. A heat-release detec-
tion method for multimodal heat-release feedback is also presented. The pro-
blem of calibrating unknown heat-release model parameters is represented as

32
 1.5 Outline and Contributions

a state-estimation problem. With this representation, a particle filter and an ex-

tended Kalman filter are evaluated for on-line estimation of cylinder-wall surface
temperature, TDC offset and a convective heat-transfer parameter.

Related Publication
Ingesson, G., L. Yin, R. Johansson, and P. Tunestål (2016). “Evaluation of nonli-
near estimation methods for calibration of a heat-release model”. SAE Int. J.
Engines 9:2, pp. 1191–1200.

Chapter 5
The setup used in the experimental work is presented in this chapter. Engine spe-
cifications, sensors and actuators are presented together with a description of the
control-system architecture.

Chapter 6
Chapter 6 investigates proportional-integral (PI) combustion-timing controller
design as a function of fuel reactivity, indicated by the fuel octane number. The
investigation was done through simulation and describes how PI controllers
should be tuned for maximized disturbance rejection subject to constraints on
robustness and cycle-to-cycle variation. The obtained results present challenges
and limiting factors for combustion-timing controller performance.

Related Publication
Ingesson, G., L. Yin, R. Johansson, and P. Tunestål (2018). “Proportional–integral
controller design for combustion-timing feedback, from n-heptane to
iso-octane in compression–ignition engines”. J. Dynamic Systems, Mea-
surement, and Control 140:5, p. 054502.

Chapter 7
This chapter covers simultaneous control of ignition delay and combustion
timing through combined actuation of fuel injection and valve positions in
the gas-exchange system. The suggested model predictive controller utilizes
a physics-based ignition-delay model, previously presented and calibrated in
Chapter 2. The controller was evaluated experimentally and was shown capa-
ble of tracking set points with respect to cylinder-individual combustion timings
and the cylinder-mean ignition delay during engine load and speed changes.

Related Publication
Ingesson, G., L. Yin, R. Johansson, and P. Tunestål (2015). “Simultaneous control
of combustion timing and ignition delay in multi-cylinder partially premixed
combustion”. SAE Int. J. Engines 8:5, pp. 2089–2098.

33
Chapter 1. Introduction

Chapter 8
Chapter 8 investigates the effects of pilot injection for control of the pressure-rise
rate. An experimental evaluation on how pilot injection affects emission levels
and efficiency is first presented. The experimental results are then used to de-
sign a model predictive controller with the objective to fulfill an upper limit on
pressure-rise rate, and to track combustion-timing and engine-load set points.
Experimental controller-evaluation results are also presented.

Related Publication
Ingesson, G., L. Yin, R. Johansson, and P. Tunestål (2016). “A double-injection
control strategy for partially premixed combustion”. In: Proc. 8th IFAC Sym-
posium on Advances in Automotive Control (AAC 2016). Vol. 49. 11. Norrkö-
ping, Sweden, pp. 353–360.

Chapter 9
Chapter 9 explores control strategies for improved combustion efficiency at
low-load operation. The results presented were based on experimental engine
data. Saturation of the fuel-injection timing and the introduction of a pilot in-
jection increased the indicated efficiency. Gas-exchange actuation for avoidance
of low-efficiency regions in a φ − T diagram was found through simulation of a
calibrated gas-exchange system model.

Related Publication
Ingesson, G., L. Yin, R. Johansson, and P. Tunestål (2016). “Control of the low-load
region in partially premixed combustion”. In: Proc. J. Physics: Conference Se-
ries. Vol. 744. 1. Southampton, England.

Chapter 10
Chapter 10 investigates potential efficiency improvements with multimodal
heat-release rates, when constraints on maximum cylinder pressure, NOx and
exhaust temperature are imposed. A simulation study showing an efficiency in-
crease with two injections is first presented. The simulation results suggest a
heuristic hybrid fuel-injection controller that varies the number of injections de-
pending on operating conditions. Experimental controller-performance results
in both steady state and transient operation are presented. The experimental
result showed a 4-5 % efficiency improvement with respect to pressure and NOx
constraints, compared to that of a single-injection controller

Related Publications
Ingesson, G., L. Yin, R. Johansson, and P. Tunestål (2017). “Efficiency optimal,
maximum-pressure control in compression-ignition engines”. In: Proc. Ame-
rican Control Conf. (ACC 2017). Seattle, WA, USA, pp. 4753–4759.

34
 1.5 Outline and Contributions

Ingesson, G., L. Yin, R. Johansson, and P. Tunestål (2018). “Design and evalua-
tion of a multiple-injection controller for efficient fulfillment of NOx and
exhaust-temperature constraints”. submitted to SAE Int. J. Engines.

Chapter 11
This chapter introduces a model-based pressure-prediction method. The novelty
of this method lies in the use of the estimated heat-release rate to predict how the
cylinder pressure varies with injection timing. This is a computationally cheap
alternative to heat-release modeling. This chapter also presents how this me-
thod can be used to find efficiency-optimal injection timings with respect to con-
straints on maximum pressure and pressure-rise rate.

Related Publication
Ingesson, G., L. Yin, R. Johansson, and P. Tunestål (2015). “A model-based
injection-timing strategy for combustion-timing control”. SAE Int. J. Engines
8:3, pp. 1012–1020.

Chapter 12
This chapter presents how the pressure-prediction method in Chapter 11 can
be used in a model predictive control framework to efficiently fulfill constraints
on in-cylinder pressure, NOx and exhaust temperature with multiple injections.
This controller objective was previously studied in Chapter 10, but was revi-
sed here with a model-based approach. An alternative model predictive con-
trol formulation for pressure set-point tracking is also presented. The control-
lers presented rely on a heat-release detection method, capable of separating the
heat-release rate among several injections. Model-based methods for adding and
removing injections are also discussed.

Related Publication
Turesson, G., L. Yin, R. Johansson, and P. Tunestål (2018). “Predictive pressure
control with multiple injections”. submitted to E-CoSM 2018, Changchun,
China.

Chapter 13
The thesis is concluded in Chapter 13, where aspects on future research are also
discussed.

Authors Contributions
The author was the foremost contributor to the publications listed above. The
author was the main contributor to the work related to modeling, control and
experimental evaluation. This includes problem formulation, as well as develo-
ping, implementing, testing, and evaluating controller designs. The author wrote

35
Chapter 1. Introduction

the papers himself with input from the co-authors. All authors jointly determined
the general direction of the research, for example, that focus should be directed
towards control problems related to fuel-injection and its effect on the combus-
tion processes. Lianhao Yin was the main contributor to the design of the expe-
rimental platform used in the thesis. This includes design of the test-cell setup,
control-system and data-acquisition architecture.

36
2
Modeling

The increasing demands for reduced fuel consumption, emissions levels and im-
proved driveability lead to more actuators, sensors and complex control func-
tions. With the increasing engine complexity in mind, a systematic implementa-
tion of the engine-control system requires mathematical models for simulation,
calibration and controller design [Isermann, 2014, Atkinsson, 2009].
Model-based control is a controller-design approach where a mathematical
model describing physical or empirical system knowledge is utilized. A model
can be used off-line to evaluate controller performance in simulation, and in this
way provide suitable controller parameters and reduce the experimental work-
load. A model can also be used on-line to provide the controller with information
about predicted system behavior in real time. Both approaches to model-based
controller design were adopted in this work.

2.1 Control-Oriented Modeling

Engine modeling can be done at different levels of detail and complexity and in-
volves several disciplines such as thermodynamics, chemical kinetics and fluid
mechanics. Computational fluid dynamics (CFD) modeling is the most com-
plex and detailed modeling approach. CFD utilizes numerical algorithms to so-
lve partial differential equations describing dynamical fluid flow with high spa-
tial and temporal resolution, whilst incorporating chemical reactions with up
to hundreds of species and reactions. This resolves details of gas dynamics,
multi-phase flows, and turbulence-chemistry interactions. Applications for CFD
modeling include design of combustion-chamber geometry, fuel-spray proper-
ties and in-cylinder flow patterns, related to engine development as an alter-
native and complement to experimental testing [Han et al., 2002; Szekely et
al., 2004; Shi et al., 2011]. CFD is also used in engine research to give physical
explanation for experimental results. However, simulation times from hours to
several months make CFD modeling unsuitable for controller design, where si-
mulation tests have to be done over a large number of engine cycles and test

37
Chapter 2. Modeling

cases. For on-line control applications, computational times have to be in the

range of milliseconds to meet the hard timing constraints of engine actuation.
Zero-dimensional (0D) models are an alternative of lower complexity. They
represent mean variables over a larger space, such as the average pressure and
temperature in the combustion-chamber volume. These models are commonly
derived from first principles, where fuel injection, fuel/air mixing and com-
bustion are represented as sub-models with empirical parameters. Short com-
putational times allow for engine simulation over complete drive cycles and
real-time applications. Examples of such models are presented in [Kiencke and
Nielsen, 2000; Isermann al., 2014; Eriksson and Nielsen, 2014]. Albeit spatial re-
solution is lost, zero-dimensional models are able to accurately predict engine
data with high temporal resolution, see [Kiencke and Nielsen, 2000; Chmela et
al., 2007; Widd et al., 2012].
Models with high temporal resolution are not always directly applicable for
controller design. Fuel-injection timings and durations are for instance only set
once every engine cycle. It would therefore be sufficient for a fuel-injection con-
troller to utilize a model describing how the engine state changes from one cycle
to the next. Cycle-to-cycle models can be obtained by sampling crank-angle re-
solved models or by deriving empirical models from engine data. Examples of
physics-based and empirical cycle-to-cycle models can be found in [Guzella and
Onder, 2009] and [Henningsson et al., 2012], respectively.
The dynamics of the fuel-delivery and gas-exchange systems have longer
time constants than the cycle-resolved combustion processes. This can be ex-
ploited by neglecting discrete engine cycles and instead model slower engine
dynamics by averaging over several engine cycles and viewing faster processes
as static relations [Eriksson and Nielsen, 2014]. These models are typically of
low spatial resolution and are referred to as mean-value models (MVM) [Hend-
ricks, 1986], or control-oriented models due to their usefulness in control appli-
cations.
The modeling detail required is a trade-off between computational complex-
ity and the possible increase in controller performance. It also depends on app-
lication requirements. For the purpose of closed-loop controller design, it is
usually sufficient to model the system behavior in the vicinity of the desired
bandwidth of the closed-loop system. This is because feedback control makes the
closed-loop system robust to model errors and disturbances of lower frequen-
cies. High-frequency components are commonly filtered out and may in those
cases be neglected by the model.
The modeling approach adopted in this work is illustrated in Fig. 2.1.
In-cylinder processes were modeled using a crank-angle-resolved zero-dim-
ensional model, derived from first principles of thermodynamics. Combus-
tion was modeled with empirical expressions for global chemical-reaction ra-
tes, triggered by the event of fuel injection. This model was occasionally sim-
plified through linearization or sampling for cycle-to-cycle controller design.

38
 2.2 In-Cylinder Model

crank-angle resolved in-cylinder 0D model

cycle k

intake → compression → expansion → exhaust

actuation for cycle k + 1

| {z }

sampling of cycle k
BC for 0D model
injection → delay → combustion

BC for MVM
MV gas-exchange model

cycle-to-cycle controller

Figure 2.1 The modeling approach adopted in this work was to represent the
in-cylinder processes with a crank-angle resolved zero-dimensional (0D) model
with empirical expressions for chemical reaction rates. Cylinder boundary condi-
tions (BC) are determined by the states of the intake and exhaust manifolds, obtai-
ned from a mean-value gas-exchange model. Controller sampling and actuation
occur in-between consecutive engine cycles.

A mean-value model was used to model states of the intake and exhaust mani-
folds which determine the boundary conditions for the in-cylinder model.
These models will be described in greater detail in the following sections.
First, the crank-angle resolved in-cylinder pressure and temperature model
is presented. The mean-value gas-exchange model is presented subsequently.
Then, ignition-delay models of different complexity are presented and evaluated.
Finally, an NOx -formation model is presented. System identification was con-
ducted in order to obtain unknown model parameters from engine data. Identi-
fication experiments and results are also presented below.

2.2 In-Cylinder Model

The main objective of this thesis is to investigate model-based combustion con-
trol though fuel-injection actuation. Models for heat-release estimation and pre-
diction of how the cylinder pressure depends on fuel injection are therefore es-
sential controller components. The models used for this purpose is based on the
assumption that the in-cylinder gas during the closed part of the engine cycle
is an open thermodynamic system with the combustion chamber as the system
boundary, see Fig. 2.2.

39
Chapter 2. Modeling

dQ ht

dQ c

combustion chamber dW

piston

Figure 2.2 Open system boundary for the combustion chamber. The combus-
tion is modeled as a release of heat dQ c , whereas the cylinder gas performs work
on the piston dW , and heat dQ ht is transferred to the cylinder walls.

The first law of thermodynamics for this system states

X
dU = dQ − dW + h i d m i (2.1)
i

where dU is the change in system internal energy, dQ is the heat added to the
P
system, dW is the work done by the system, and i h i d m i is the enthalpy flux
across the system boundary. The combustion is modeled as a release of heat,
which gives that the heat added to the system dQ is the difference between rele-
ased chemical energy dQ c and heat transferred to the cylinder walls

dQ = dQ c − dQ ht (2.2)

With the piston work given by pdV , the first law can be rewritten as
X
dU = dQ c − dQ ht − pdV + h i d m i (2.3)
i

Under the assumption of ideal gas, a change in internal energy is given by

dU = mc v d T + ud m (2.4)

where m is the system mass, c v is the heat-capacity at constant volume and u is

the specific internal energy. The ideal-gas law, where R is the specific gas con-
stant
pV = mRT (2.5)
provides a differential in T
1
dT = (V d p + pdV − RT d m) (2.6)
mR

40
 2.2 In-Cylinder Model

by assuming a constant R. Now, by inserting (2.4) and (2.6) into (2.3), and substi-
tuting
R
cv = (2.7)
γ−1

where γ is the ratio of specific heats γ = c p /c v , the following differential in p can

be derived

γ γ−1 X 1
dp = − pdV + (dQ c − dQ ht + h i d m i ) + (RT − (γ − 1)u)d m (2.8)
V V i V

From (2.8), the simplified differential equation in p is obtained

µ ¶
dp γ dV γ − 1 dQ c dQ ht
=− p+ − (2.9)
dθ V dθ V dθ dθ

if m is assumed to be constant, and the crank angle θ is chosen as the indepen-

dent variable. In this work, (2.9) was used to analyze dQ c /d θ from measured p.
A more detailed description of heat-release analysis is given in Chapter 4. Model
(2.9) was also used to simulate p with initial conditions and dQ c /d θ as input in
Chapter 10.
The assumption of a uniform ideal gas with constant mass and composition
is a useful approximation that is adequate for many engineering applications.
In reality, liquid fuel is injected into the cylinder. The fuel vaporizes and mixes
with air to produce a nonuniform fuel and air distribution, where the chemical
composition changes during combustion. Flows to and from cooling crevice re-
gions have a significant effect but was also neglected here. More detailed models
that take into account for mass flow to crevice volumes, radiation, and thermal
boundary layers are presented in [Heywood, 1988]. Models that divide the com-
bustion chamber into multiple zones, instead of one, can be found in [Fiveland
and Assanis, 2001; Nilsson and Eriksson, 2001].
The following sections describe how the different components of (2.9) were
modeled.

Cylinder Geometry
The cylinder volume V was modeled as a slider-crank mechanism
µ r ¶
Vd π π
V = Vc + R v + 1 − cos( θ) − R v2 − sin2 ( θ) (2.10)
2 180 180

where Vd and Vc are displacement and clearance volumes, and R v is the ratio of
the connecting-rod length to the crank radius.

41
Chapter 2. Modeling

Temperature
The in-cylinder temperature T was computed from the ideal-gas equation using
the conditions at intake-valve closing (IVC)

pV TIVC
T= (2.11)
p IVC VIVC

The temperature TIVC was computed as a weighted average of the intake-manifold

gas temperature Tin and the trapped residual-gas temperature Tres

c vin Tin + c vres αχTres

TIVC = (2.12)
c vin + c vres αχ

where c vin and c vres are specific heats of inducted charge and residual gases, α is
the mass ratio between trapped residuals and inducted gases and χ is a measure
of residual-gas temperature decrease during gas exchange. The residual gas tem-
perature Tres was computed from the exhaust-valve-opening temperature of the
previous cycle.

Ratio of Specific Heats

The ratio of specific heats γ = c p /c v is determined by the cylinder-gas composi-
tion and temperature. In this work, γ was held constant for less sensitive compu-
tations, such as determining the combustion-timing. The ratio γ was computed
with NASA specific-heat polynomials as a function of in-cylinder gas tempera-
ture and composition when an increased model accuracy was needed.
The gas composition is non-trivial to compute during combustion and was
therefore interpolated between the unburned and burned gas composition du-
ring combustion, using the computed heat-release rate, EGR ratio, and the stoi-
chiometry of the overall chemical reaction

1 ³ y´
Cx H y + x+ (O2 + 3.773N2 ) →
φn 4
µ ¶
y 1 ³ y´ 1 ³ y´
xCO2 + H2 O + − x+ O2 + 3.773 x + N2 (2.13)
2 φn 4 φn 4

where y/x is the fuel hydrogen to carbon ratio and φn is the molar equivalence
ratio.

Heat-Transfer Model
The convective heat-transfer rate from in-cylinder gas to piston, cylinder head
and walls dQ ht /d θ was modeled using Newton’s law of cooling

dQ ht hc A
= (T − T w ) (2.14)
dθ 60Nspeed

42
 2.2 In-Cylinder Model

where T w is the wall-surface temperature, A is the combustion-chamber surface

area and Nspeed is engine speed in revolutions per minute (rpm). The convection
coefficient h c is an empirical [Woschni, 1967] expression

h c = 3.26B 0.2 p 0.8 T −0.55 ω0.8 (2.15)

derived from a Nusselt-Reynolds correlation, as presented in [Heywood, 1988].

Here, B is the cylinder bore and ω the mean cylinder-gas velocity, where ω was
modeled according to

Vd TIVC
ω = C 1 S p +C 2 (p − p m ) (2.16)
p IVC VIVC

The first term in (2.16) relates gas motion to the mean piston speed S p , and the
second term captures the effect of charge-density variation during combustion,
where p m is the cylinder pressure of a motored cycle. The empirical parame-
ters C 1 and C 2 are engine dependent, where C 2 = 0 before the start of combus-
tion. The problem of estimating C 2 from cylinder-pressure data is investigated in
Chapter 4. Similar global heat-transfer models presented in [Annand, 1963; Ho-
henberg, 1979] are also commonly used to model (2.14).

Cylinder-Wall Temperature
The cylinder-wall surface temperature T w is determined by the heat flux from
in-cylinder gases to the engine coolant with temperature Tc . The cylinder wall,
cylinder head and piston were modeled as a single mass with conductive coeffi-
cient k c , thickness L c , mass m c and specific heat c p . By assuming that the inner
wall temperature is in steady state and that the outer-wall surface on the coo-
lant side has fixed temperature Tc , the following dynamic equation for T w can be
derived
d Tw 2A(h c + k c /L c ) 2Ah c
=− Tw + (T + Tc ) (2.17)
dθ m c c p 60Nspeed m c c p 60Nspeed

During the intake and exhaust strokes, T was assumed to be constant and equal
to intake- and exhaust-manifold temperatures. This model was previously pre-
sented and used in [Roelle et al., 2006; Widd et al., 2008].

Heat-Release Rate
The heat-release rate dQ c /d θ is difficult to model from first principles due
to its dependency on a multitude of factors such as chemical combus-
tion rates, fuel-injection profile and fuel-air mixing rates. The approach ad-
opted in Chapter 11 was to utilize the heat-release rate computed from
cylinder-pressure measurements to predict how p varies with fuel injection.
This method can not be used to simulate p. A Wiebe [1970] expression for the

43
Chapter 2. Modeling

accumulated heat-release
 µ ¶
 ³ θ − θSOC b+1 ´
Q c (θ) 1 − exp − a for θ ≥ θSOC
= ∆θ (2.18)
Q tot 
0 otherwise
was therefore used, mainly because of its simplicity, as an input to simulate (2.9)
in Chapter 10. In (2.18), θSOC is the start of combustion, and the parameters ∆θ,
a and b relate to the duration and shape of the heat-release profile.

Load and Efficiency Definitions

The measured in-cylinder pressure was used to compute the engine work out-
put and efficiency. The gross- and net-indicated mean effective pressures, p IMEPg
and p IMEPn are normalized measures of the work done on the piston by the cylin-
der gas during the closed part of the cycle and during the complete cycle. These
quantities were here defined as
Z
1 EVO
p IMEPg = pdV
Vd IVC
I (2.19)
1
p IMEPn = pdV
Vd
The gross mean effective pressure p IMEPg is more commonly defined by
Z
1 BDC2
p IMEPg = pdV (2.20)
Vd BDC1
where the integral is taken over the complete compression and expansion stro-
kes. The difference between (2.19) and (2.20) is determined by engine valve ti-
mings and intake and exhaust manifold pressures. The reason for using the defi-
nition in (2.19) was to distinguish the efficiency of the closed part of the cycle and
to evaluate efficiency in simulation without a gas-exchange model. Both p IMEPg
and p IMEPn were used as feedback variables to control the engine load. For this
purpose, the difference between (2.19) and (2.20) has negligible effect.
With (2.19) and (2.20) computed, fuel-conversion efficiencies to indicated
work are given by
p IMEPg Vd
η GIE =
m f Q LHV
(2.21)
p IMEPn Vd
η NIE =
m f Q LHV
where m f is the injected fuel mass and Q LHV is the lower heating value of the
fuel. These efficiencies were frequently used to evaluate simulated and measured
engine efficiency. The gross efficiency η GIE is of interest if the combustion and
thermodynamic efficiencies of the closed part of the engine cycle are of interest,
whereas η NIE also accounts for pumping losses during gas exchange.

44
 2.3 Gas-Exchange System

air

intercooler θcool

1
θhot EGR

3 4 EGR

compressor
θLP
θHP

turbine
exhaust manifold
intake manifold

θBP
- valve

exhaust

Figure 2.3 Engine gas-exchange system layout. The engine was boosted by a
fixed-geometry turbocharger. Exhaust-gas recirculation was supplied by a high
and a low-pressure path, and the intake temperature was controlled by the gas
flow through a charge-air cooler prior to the intake manifold. The gas-exchange
system was modeled as 5 interconnected adiabatic volumes (indicated by the blue
regions). Gas flows through the EGR paths and the thermal management system
were controlled by actuating valve positions, denoted θx . The engine layout is fur-
ther described in Chapter 5.

2.3 Gas-Exchange System

A layout of the gas-exchange system used in the experimental work is presented
in Fig. 2.3. Air enters from above and is compressed by a turbocharger. The air

45
Chapter 2. Modeling

then either goes through a charge-air cooler or directly to the intake manifold.
The mass-flow ratio between these paths is determined by the position of two
valves. These paths will later be referred to as the fast thermal-management
part (FTM) of the gas-exchange system. After leaving the exhaust manifold,
the exhaust gases either enter the cooled high-pressure EGR path or expands
through the turbine. After the turbine, some of the exhaust goes through a co-
oled low-pressure EGR path, and the remaining exhaust gas passes through a
back-pressure valve to the exhaust pipe. Mass flows through the EGR paths are
determined by the EGR-valve positions.

Gas-Exchange Dynamics
A gas-exchange system model was used in Chapter 9 to design a low-load PPC
controller. The main objective of the model was to describe in-cylinder tempe-
rature and mass in the low-load operating range of the engine, in order to com-
pute optimal valve-position actuation in simulation. The gas-exchange system
was modeled as five adiabatic ideal-gas control volumes, interconnected with
restrictions, denoted 1 to 5 in Fig. 2.3. Under this assumption, the dynamic equa-
tions with respect to pressure p i and temperature Ti in volume Vi are given by
[Eriksson and Nielsen, 2014]

d p i RTi p i d Ti
= (ṁ in − ṁ out ) +
dt Vi Ti d t
(2.22)
d Ti RTi
= (ṁ in c v,i (Tin − Ti ) + R(ṁ in Tin − ṁ out Ti ))
dt Vi p i c v,i
Here, ṁ in and ṁ out are in- and outgoing mass flows, and c v,i is the gas speci-
fic heat at constant volume. States in the intake and exhaust manifolds are also
denoted X in and X ex in this thesis. Mass flow from Vi to V j , ṁ i j , was mode-
led as turbulent compressible flow through a restriction, where ṁ i j is given by
[Heywood, 1988]

µ ¶ s µ µ ¶¶(γi −1)/γi
A i j p i p j 1/γi 2γi pj
ṁ i j =p 1− (2.23)
RTi p i γi − 1 pi
Here, A i j is the effective flow area and γi is the ratio of specific heats which is
different for air and exhaust. If the pressure ratio is too low

µ ¶ γi
pj 2 γi − 1
< (2.24)
pi γi + 1
the flow becomes choked, and the expression for ṁ i j is given by

µ ¶ γi + 1
A i j p i 1/2 2 2(γi − 1)
ṁ i j =p γ (2.25)
RTi i γi + 1

46
 2.3 Gas-Exchange System

The effective flow area A i j was in the case of a valve restriction determined by
the valve angle θi j according to an empirical expression
³ ´
−k (θ −θ 0 )
A i j = A 0i j 1 − e Ai j i j i j (2.26)

in
The mass flow from the intake manifold to the cylinders ṁ cyl was modeled using
a volumetric efficiency η v , the intake-manifold density and the engine displace-
ment rate
in p 3 Vd Nspeed
ṁ cyl = ηv · · (2.27)
RT3 120
The volumetric efficiency was approximated by assuming displacement of new
charge by residual gas
µ ¶1/γ
p4
rc −
p3
ηv = (2.28)
rc − 1
where r c is the engine compression ratio. The cylinder-out mass flow was then
obtained by adding the fuel flow ṁ f

out in
ṁ cyl = ṁ cyl + ṁ f (2.29)

The turbocharger model used was a simplification of the model presented

in [Wahlström and Eriksson, 2011]. First, the following relation for the turbine
power P t
P t = η t P ts (2.30)
was assumed where η t is the turbine thermodynamic efficiency and P ts is the
power obtained from isentropic expansion over the turbine
à µ ¶1−1/γ4 !
s p5
P t = c p,4 T4 1 − ṁ 45 (2.31)
p4

where ṁ 45 is the mass flow over the turbine, modeled using (2.23). The compres-
sor power P c was then related to P t through a mechanical efficiency η m and the
static relation
P c = ηm P t (2.32)
The compressor power P c was in turn related to the power of isentropic compres-
sion over the compressor P cs
P cs = η c P c (2.33)
where P cs is a function of the compressor mass flow ṁ 12
õ ¶ !
s p 2 1−1/γ1
P c = c p,1 T1 − 1 ṁ 12 (2.34)
p1

47
Chapter 2. Modeling

We can now solve for ṁ 12 which gives

à ¶µ !
p 5 1−1/γ4
η tc c p,4 T4 1 −
p4
ṁ 12 = õ ¶ ! ṁ 45 (2.35)
p 2 1−1/γ1
c p,1 T1 −1
p1

where η tc is the overall efficiency of the turbocharger

η tc = η c η m η t (2.36)

More detailed turbocharger models include dynamics and experimentally obtai-

ned maps describing turbine and compressor efficiencies. The approach taken
here was instead to calibrate a constant η tc over a limited engine operating range.
Engine-out temperature TEO and cooler-out flow temperatures TEGR , TCO were
modeled using empirical polynomials
TEO (T3 , m f ) = c EO (T3 , m f )
(2.37)
ij ij
TCO (Ti , ṁ i j ) = c CO (Ti , ṁ i j )
The unknown model parameters were identified with respect to engine data
by first calibrating the polynomial coefficients in c X and k A i j in (2.26) with re-
spect to local flow and temperature data. The remaining unknown parameters
¡ ¢T
ϑ = A atm,1 A 123 A 223 A 43 A 45 A 51 A 5,atm η tc (2.38)

were then computed by minimizing the model-output error cost function

5
X
V (ϑ) = αi kp im − p i k22 + βi kTim − Ti k22 (2.39)
i =1

where p im , Tim are measured pressure and temperature data, and p i , Ti are cor-
responding model outputs. The importance of capturing different outputs are
determined by the tuning parameters αi and βi . Cost function (2.39) was mini-
mized subject to
η tc ≤ 1 (2.40)
and measured boundary conditions p atm , Tatm . Model output and measured
data are shown in Fig. 2.4 for a local minimizer of (2.39) ϑ∗ , where ϑ∗ was
found using the MATLAB nonlinear-optimization toolbox. The data presented
in Fig. 2.4 are combined sequences of steady-state data with different θCool and
θHP positions for p IMEP between 1 and 5 bar. The tuning parameters were set to
prioritize model fit with respect to the intake-manifold state. In Fig. 2.4, it can be
seen that the model managed to predict in-cylinder temperature at the start of
injection TθSOI and the relative air-fuel ratio λ, which was the main purpose of
the model when used in Chapter 9.

48
 2.3 Gas-Exchange System

1.2 1.6
1.4

p ex [bar]
p in [bar]
1.1
1.2
1
1
0.9
0 500 1,000 1,500 2,000 0 500 1,000 1,500 2,000
cycle cycle

340 800

TθSOI [K]
320 750
Tin [K]

300 700

280 650
0 500 1,000 1,500 2,000 0 500 1,000 1,500 2,000
cycle cycle

1.4
1.3 40
θ HP [deg]
p 2 [bar]

1.2
20
1.1
1 0
0 500 1,000 1,500 2,000 0 500 1,000 1,500 2,000
cycle cycle

100 20
80
θ cool [deg]

15
60
λ [-]

10
40
20 5
0 0
0 500 1,000 1,500 2,000 0 500 1,000 1,500 2,000
cycle cycle

Engine Data Model Output

Figure 2.4 Model output (black) and measured data (blue) for a local minimizer
ϑ∗ of (2.39). The light blue color indicates cycle-to-cycle variation in the intake
and exhaust manifolds. The engine data presented are combined sequences of
steady-state data with different θCool and θEGR positions for p IMEPn from 1 to 5
bar. The model managed to capture in-cylinder temperature at the start of injec-
tion TθSOI and the relative air-fuel ratio λ.

49
Chapter 2. Modeling

Cylinder Oxygen Concentration

The EGR mass flow reduces the cylinder oxygen concentration, which in turn af-
fects chemical reaction rates and the ignition delay. The oxygen concentration
was for this reason an important ignition-delay model input and was computed
from air-flow measurements and estimated EGR mass flow according to the fol-
lowing description. The oxygen concentration at intake-valve closing [O2 ]IVC is
determined by the oxygen in the inducted air, EGR, and residual gas
µ ¶
120ṁ air 120ṁ EGR m res
[O2 ]IVC = χair
O2
ex
+ χO2 + (2.41)
6M air Vd Nspeed 6M ex Vd Nspeed M ex Vd

where ṁ air is the measured engine air flow, M air , M ex , χair ex

O2 and χO2 are molar
masses and oxygen molar fractions of air and exhaust. The latter was computed
from λ-sensor measurements. A Kalman filter [Kalman, 1960] (see Chapter 3) was
then used to estimate the EGR mass flow from both EGR paths ṁ EGR with the use
of the intake-manifold pressure model

d p in RTin in p in d Tin
= (ṁ air + ṁ EGR − ṁ cyl )+ (2.42)
dt Vin Tin d t

A similar EGR-estimation technique was presented in [Lee et al., 2013]. The

residual-gas mass m EGR was computed using the state in the exhaust manifold
at exhaust valve opening (EVO)

p ex VEVO
m res = (2.43)
RTex

This method was used to compute [O2 ]IVC online in Chapter 7, where [O2 ]IVC was
used to predict ignition-delay variation.

2.4 Ignition-Delay Modeling

The ignition delay is an important variable in PPC where it should be kept suf-
ficiently long for the fuel to mix with the cylinder-gas mixture before the start of
combustion. In a direct-injection combustion engine, the ignition delay depends
on physical processes such as fuel atomization, vaporization and the mixing of
fuel and air in the cylinder. It also depends on chemically controlled autoignition
reactions [Heywood, 1988]. Here, the ignition delay τ is defined as the time in
milliseconds between θSOI and the crank angle of 10 % heat released θ10 , which
was used as an indicator for ignition, see Fig. 2.5

θ10 − θSOI
τ= (2.44)
0.006Nspeed

50
 2.4 Ignition-Delay Modeling

6,000 p

Qc
Q c [J] 4,000

2,000

−2,000 θ SOI τ θ10

−20 −15 −10 −5 0 5 10

θ [CAD]

Figure 2.5 The ignition delay τ was defined as the time between the start of in-
jection θSOI and the crank angle of 10 % heat released θ10 . This indicator for the
start of combustion was computed from the accumulated heat release Q c . The
start of injection was here determined by the rising flank of the injector-current
pulse and a hydraulic injector delay.

For the purpose of controller design, autoignition models that correlate τ

with the cylinder-gas state were studied under the assumption that the chemi-
cally controlled part of τ contributes most to its variability. This assumption is va-
lid for longer τ where the relative importance of autoignition is high. The relative
importance of physical factors becomes more important for shorter τ [Heywood,
1988].
Detailed reaction mechanisms for larger hydrocarbons consist of thousands
of species and tens of thousands of reactions [Lu and Law, 2009]. Model complex-
ity can be reduced by removing or lumping reactions of less importance, or by
building empirical models that uses a sufficient number of species and reactions
to capture experimental τ data. A well-known example of the latter is the global
reaction mechanism with five representative species and eight reactions deve-
loped by Halstead et al. [1975, 1977]. This model was able to accurately predict
two-stage ignition of alkanes under engine-like conditions in a rapid compres-
sion machine.
The empirical modeling approach was adopted in this work, where three
different low-order models denoted M1, M2 and M3 with increasing complex-
ity were evaluated. An additional model, M4, derived from chemical-reaction
simulations by Delvescovo et al. [2016] was used to simulate τ for different
octane numbers, for the purpose of evaluating the fuel-reactivity effect on
combustion-timing controller design. A reason for using simpler ignition-delay
models was the need to simulate in the order of 107 engine cycles when evalua-
ting different controller parameters in Chapter 6. Another reason was the require-
ment for computational times below the engine-cycle duration when regulating
τ in Chapter 7. Models M1 to M4 are presented in the following sections.

51
Chapter 2. Modeling

Evaluation of M1, M2 and M3

The three low-order autoignition models, M1, M2, and M3, were evaluated for
the purpose of model-based control. The evaluation was done by studying how
well these models, when calibrated, could predict experimental τ data for dif-
ferent injection timings, intake temperatures and oxygen concentrations, in the
low-to-mid load and speed range of the engine. The complexity of linearizing the
models for controller design was also investigated. Another reason for evaluation
the models was to gain a better understanding of the model complexity needed
to predict experimental data.
These models are based on the fact that the reaction rate for a chemical re-
action with two species, B and C , is proportional to the concentrations of the
reactants via a rate coefficient k reac
d [B ]
= −k reac [B ][C ] (2.45)
dt
The rate coefficient is usually modeled with a temperature dependent Arrhenius
expression
k reac = Ae −E a /R̃T (2.46)
where A is a collision-frequency constant, R̃ is the universal gas constant, and
E a is an activation energy. Although combustion proceeds in multiple steps, a
useful approximation can be obtained by modeling the reaction rate as a single
mechanism [Turns, 1996]
d [C x H y ]
= −A[C x H y ]a [O 2 ]b e −E a /R̃T (2.47)
dt
where [C x H y ] and [O 2 ] are fuel and oxygen concentrations, and a and b are em-
pirical parameters.

M1
The first model is given by the following inverted reaction-rate expression
α β
τ = A[O2 ] [Cx Hy ] p̄ ζ S δp e E a /R̃ T̄ (2.48)

where A, E a , α, β, ζ and δ are fuel-dependent empirical parameters, and the no-

tation X̄ denote the average X between θSOI and θ10 , under the assumption that
all fuel has been injected. The unknown model parameters

ϑM 1 = [A, E a , α, β, ζ, δ]T (2.49)

are to be identified from engine data.

Similar models were presented by [Heywood, 1988] and were used by Kem-
pinski and Rife [1981] and also by Donald and Eyzat [1978] to parameterize τ for
different fuels and engines.

52
 2.4 Ignition-Delay Modeling

M2
The second model is a crank-angle resolved extension of M1 where it is assumed
that combustion starts when
Z θSOI +τ
A[O2 ]α [Cx Hy ]β p ζ S δp e −E a /R̃T d θ = 1 (2.50)
θSOI

is fulfilled. The integral relation can be interpreted as an autoignition criterion,

where combustion is initiated when a critical number of radicals is reached
[Chiang and Stefanopoulou, 2009]. The unknown model parameters are here
given by
ϑM 2 = [A, E a , α, β, ζ, δ]T (2.51)
This model was previously presented by Thurns [1996] and Chmela et al. [2007]
among others.

M3
The third model is a two-stage reaction model with empirically determined reac-
tion rates as presented in [Westbrook and Dryer, 1981]
¡x
y ¢ k1 y
Cx Hy + O2 → xCO + H2 O
+
2 4 2
(2.52)
1 k2
CO + O2 → CO2
2
Here, the fuel Cx Hy reacts with O2 to form CO and H2 O in the first reaction,
CO is then oxidized to CO2 in the second reaction. Reaction rates are given by the
following differential equations

d [Cx Hy ]
= −k 1 [Cx Hy ]β [O2 ]α
dt
d [O2 ] ¡x y ¢
= − + k 1 [Cx Hy ]β [O2 ]α − k 2 [CO][O2 ]1/2
dt 2 4
(2.53)
d [CO]
= xk 1 [Cx Hy ]β [O2 ]α − k 2 [CO][O2 ]1/2
dt
d [CO2 ]
= k 2 [CO][O2 ]1/2
dt
Initial conditions for (2.53) are given by the global cylinder oxygen and fuel
concentrations after fuel injection. The reaction-rate parameters k i are empirical
expressions on the form
i
k i = A i S pi e −E a /R̃T
δ
(2.54)
Finally, the heat-release rate of the reactions is computed from
dQ c
= V (Q 1 k 1 [Cx Hy ]β [O2 ]α + k 2Q 2 [CO][O2 ]1/2 ) (2.55)
dt

53
Chapter 2. Modeling

Table 2.1 Investigated operating points for model calibration and evaluation.

L1S1 L2S1 L1S2 L2S2

Nspeed [rpm] 1200 1200 1500 1500
p IMEPg [bar] 5 10 5 10
Tin [◦ C] 20 - 40 22 - 60 22 - 50 38 - 85
[O2 ]IVC [mol/m3 ] 7.5 - 9.5 9.5 - 11.5 7.5 - 10 8.5 - 13
θSOI [CAD] -25 - (-5) -10 - 2 -25 - 5 -15 - 0

where Q 1 and Q 2 are the lower heating values per mole of fuel and CO in (2.53).
Combustion is assumed to start when Q c has reached 10 % of the expected total
heat released Q ctot . For M3, unknown model parameters are given by

ϑM 3 = [A 1 , A 2 , E a1 , E a2 , α, β, δ1 , δ2 ]T (2.56)

Ignition-Delay Experiments
Ignition-delay experiments were conducted at the four operating points obtai-
ned by combining p IMEPg = 5, 10 bar and Nspeed = 1200, 1500 rpm, see Table 2.1.
Suitable θSOI were found at each operating point with EGR valves closed and
both thermal-management valves opened at 45◦ . A layout of the experimen-
tal engine setup showing valve locations was presented in the section covering
gas-exchange modeling above, see Fig. 2.3.
In order to investigate the τ response to different engine inputs, θSOI ,
thermal-management (θcool/hot ) and EGR-valve positions (θHP/LP ) were varied
manually in steps during a total of 12000 cycles at each operating point.
The thermal-management valve positions were changed by setting

θcool = cos−1 (1 − cos(θhot )) (2.57)

and varying θhot in order to keep an approximately constant total valve-opening

area. The low-pressure EGR path was used to adjust EGR flow at the higher load
operating points while the high-pressure EGR path was used at the lower load
operating points. The reason for doing so was an insufficient pressure diffe-
rence over the high-pressure EGR valve at higher loads. The resulting data can
be viewed in Figs. 2.6-2.9 where Tin , [O2 ]IVC , θSOI and τ for one of the six cylin-
ders are presented.
The first half of the cycles τID were used for model calibration and the
other half τVAL for model validation. This approach of evaluating model per-
formance is called cross validation and is a technique for evaluating how well
parameter-identification results generalize to an independent data set. This is
done to avoid overfitting with respect to the identification data set [Geisser,
1993].

54
 2.4 Ignition-Delay Modeling

Ignition Delay Intake Temperature

50

3
τID τVAL 40

Tin [◦ C]
τ [ms]

30
2
20

1 10
0 0.5 1 0 0.5 1
cycle [-] ·104 cycle [-] ·104

Injection Timing Oxygen Concentration

0

[O2 ]IVC [mol/m3 ]

9
θ SOI [CAD]

−10

8
−20

−30 7
0 0.5 1 0 0.5 1
cycle [-] ·104 cycle [-] ·104

Figure 2.6 Experiment data at operating point L1S1, top left: τ [ms], top right:
Tin [C ], bottom left: θSOI [CAD ATDC] and bottom right: [O2 ]IVC [mol/m3 ].

Parameter Identification
Sum-of-squares model-error cost functions J M x (τID , ϑM x ) were minimized for
each model with respect to ϑM 1−3 and the identification data set τID to identify
suitable model parameters.

M1 If logarithms are applied to both sides of (2.48), the problem of minimizing

the sum of squares

N ¡
X ¢2
J M 1 (τID , ϑM 1 ) = ln(τID M1
i ) − ln(τi ) (2.58)
i =1

with respect to ϑM 1 is a linear regression, where the index i denotes sample num-
ber and N is the number of samples. In the following notation

¡ ¢T
y = ln(τ1M 1 ) ... ln(τM 1
N ) (2.59)

55
Chapter 2. Modeling

Ignition Delay Intake Temperature

1.6
τID τVAL 60
1.4

Tin [◦ C]
τ [ms]

1.2
40
1

0.8 20
0 0.5 1 0 0.5 1
cycle [-] ·104 cycle [-] ·104

Injection Timing Oxygen Concentration

5 12

[O2 ]IVC [mol/m3 ]

0
θ SOI [CAD]

11
−5
10
−10

−15 9
0 0.5 1 0 0.5 1
cycle [-] ·104 cycle [-] ·104

Figure 2.7 Experiment data at operating point L2S1.

 
1
1 R̃ T̄1
ln([O2 ]1 ) ln([Cx Hy ]1 ) ln(p̄ 1 ) ln(S p 1 ) 
 
. .. .. .. .. .. 
Φ =  .. . . . . .  (2.60)
 
 1 
1 ln([O2 ]N ) ln([Cx Hy ]N ) ln(p̄ N ) ln(S p N )
R̃ T̄ N
¡ ¢T
the minimizer of (2.58), ϑ∗ = ln(A ∗ ) E a∗ α∗ β∗ ζ∗ δ∗ , is given by

ϑ∗ = (ΦT Φ)−1 ΦT y (2.61)

Linear-regression problems and the least-squares method are described in grea-

ter detail in [Johansson, 1993].

M2 For the second model, parameters were found by minimizing

à !2
N Z
X θSOI +τID
i
ID β ζ δ −E a /(R̃Ti )
J M 2 (τ , ϑM 2 ) = A[O2 ]α
i [Cx Hy ]i p i S p,i e dt −1 (2.62)
i =1 θSOI

56
 2.4 Ignition-Delay Modeling

Ignition Delay Intake Temperature

3 50

2.5 τID τVAL 40

Tin [◦ C]
τ [ms]

2 30

1.5 20

1 10
0 0.5 1 0 0.5 1
cycle [-] ·104 cycle [-] ·104

Injection Timing Oxygen Concentration

0 10

[O2 ]IVC [mol/m3 ]

θ SOI [CAD]

9
−10

8
−20

7
0 0.5 1 0 0.5 1
cycle [-] ·104 cycle [-] ·104

Figure 2.8 Experiment data at operating point L1S2.

This is not a linear regression, without a closed-form expression for the minimi-
zer. The cost function was therefore minimized numerically, using the MATLAB
Optimization Toolbox. In order to avoid local minima, (2.62) was minimized with
respect to different initial parameters.

M3 The reaction-rate parameters of the third model were found using the same
numerical procedure as for M2. The chemical reactions were simulated by app-
roximating the derivatives in (2.53) using the forward-Euler method and a suffi-
ciently small step size, h = 0.01 ms. The model parameters were then found by
minimizing
N ³
X ´2
J M 3 (τID , ϑM 3 ) = M 3 ID
Q c,i (θ10,i ) − 0.1m f ,i Q LHV (2.63)
i =1

M 3 VAL
where Q c,i (θ10,i ) is the modeled accumulated heat-release at θ10 in the
identification-data set, and m f ,i Q LHV is the injected fuel energy, computed from
fuel-flow measurements.

57
Chapter 2. Modeling

Ignition Delay Intake Temperature

100
1.4
80
1.2 τID

Tin [◦ C]
τ [ms]

60
1
40
0.8
20
0 0.5 1 0 0.5 1
cycle [-] ·104 cycle [-] ·104

Injection Timing Oxygen Concentration

20 14

[O2 ]IVC [mol/m3 ]

0 12
θ SOI [CAD]

−20 10

−40 8

−60 6
0 0.5 1 0 0.5 1
cycle [-] ·104 cycle [-] ·104

Figure 2.9 Experiment data at operating point L2S2. Unfortunately, the last 2000
cycles of τID were corrupted due to a malfunctioning thermal-management valve.
This part of τID was therefore replaced with the last 2000 cycles of τVAL . Note the
unintentional decrease in θSOI after cycle 11000. It was decided to be keep these
data points in the sets τID and τVAL .

Model Evaluation
The three models were evaluated by how well they could explain variation in the
validation-data set τVAL in two different ways:

• By how well a model calibrated by τID from one operating point could pre-
dict τVAL from the same operating point. During model calibration, the
speed dependence was removed (δ = 0). The load dependence was also
removed by setting β = 0 in M1-2 and β = 1 in M3. Furthermore, M1 and
M2 were investigated with and without pressure dependence, i.e., with ζ
free and ζ = 0.

• By how well a model calibrated with the complete τID data set could pre-
dict the complete τVAL data set. Now, β and δ were free parameters during
model calibration.

58
 2.4 Ignition-Delay Modeling

L1S1 L1S2
0.15 0.15
RMSE [ms] without p with p

RMSE [ms]
0.1 0.1

0.05 0.05

M1 M2 M3 M1 M2 M3

L2S1 L2S2
0.15 0.15
RMSE [ms]

RMSE [ms]
0.1 0.1

0.05 0.05

M1 M2 M3 M1 M2 M3

Figure 2.10 Model RMSE with respect to validation data for M1, M2 and M3 at
each operating point. Including p dependence in M1 and M2 did not give any
improvement. Increasing model complexity from M1 to M2 and M3 did not ne-
cessarily yield a smaller prediction error.

The reason for evaluating the models with respect to these two aspects was to de-
termine if there is an incentive to use multiple models with fewer parameters, in-
stead of having one model that covers all operating points with more parameters.
Model performance was evaluated by computing the root-mean-square error
v
u N
u1 X
RMSEM x = t (τM x − τVAL )2 (2.64)
N i =1 i i

where τM x is the model output and τVAL is measured τ in the validation-data set.

Individual Operating Points

Model RMSE scores with respect to individual operating points are presented in
Fig. 2.10. The blue and red bars indicate RMSE with and without p dependence.
It can be seen that including p did not improve performance significantly. It
rather increased error slightly, which could be a sign of overparameterization. In-
creasing model complexity from M1 to M2 decreased performance at L2S1 and

59
Chapter 2. Modeling

Table 2.2 Complexity in number of arithmetic operations where addition, sub-

traction and multiplication are all considered to be one operation. N is the num-
ber of samples between θSOI and θ10 , and τ/h is the ratio of τ and the step size
used in the forward-Euler approximation. From the author’s experience, h should
be smaller than 0.05 ms. * = crank angle resolution of 0.2 CAD, and τ = 10 CAD. **
= τ/h = 300.
Model Arithmetic Operations Example
M1 7N 750∗
M2 43N 2150∗
M3 105τ/h 31500∗∗

L2S2 but not at L1S1 and L1S2. Increasing model complexity from M2 to M3 im-
proved model performance in all operating points but L2S2.
For a more detailed analysis, the model outputs are displayed together with
τVAL in Figs. 2.11-2.14, where the gray dots indicate τVAL and the red, blue and
green lines are model outputs from M1, M2 and M3 without pressure depen-
dence.
In Fig. 2.11, the model outputs did not follow τVAL well around cycles 100 and
1000. This occured when θSOI was delayed at cycles 6100 and 7000 in Fig. 2.8. The
same behavior is found in Fig. 2.12 from cycle 1 to 800, in Fig. 2.13 at cycle 200,
and in Fig. 2.14 at cycles 500 and 1500. As θSOI was delayed close to TDC, the
computed T during τ increased which caused the models to predict a decreased
τ. Instead, the measured τVAL increased. Hence, the models were incapable of
anticipating the point where a delayed θSOI started to increase τ. This behavior
can clearly be seen in Fig. 2.12 at cycle 1400, where an increase in θSOI resulted in
a large increase in τVAL . The models were overall better at predicting τVAL during
variations in TIVC , [O2 ]IVC and θSOI for early θSOI , see cycle 750 in Fig. 2.11.
Similar model outputs for M1, M2 and M3 in Figs. 2.11-2.14 indicate that
controller designs based on the different models would yield comparable per-
formance.

All operating points

The models were also calibrated with the complete τID data set to see how well
one model could predict the overall τ behavior. The parameters β and δ were free
parameters during identification for model M1-2, and β and δ1 were free para-
meters during identification for M3. The RMSE scores are presented in Fig. 2.15,
where the green bars indicate total RMSE for the individual operating-point mo-
dels in Fig. 2.10 for comparison.
Increasing model complexity from M1 to M3 resulted in an increased RMSE,
and using operating-point individual models gave a slight reduction in RMSE.
Modeled τ for all operating points is presented together with τVAL in Fig. 2.16.

60
 2.4 Ignition-Delay Modeling

Model-Prediction Performance L1S1

3.2
τVAL
3 M1
M2
2.8 M3
τ [ms] Cross-Validation Data

2.6

2.4

2.2

1.8

1.6

1.4
0 1,000 2,000 3,000 4,000 5,000 6,000
cycle [-]

Figure 2.11 Modeled τ together with τVAL from operating point L1S1. It can be
noted that the models did not follow τVAL well at cycles 100 and 1000. M1 had
overall comparable performance to M2 and M3.

Linearization for Control Purposes

For the purpose of model-based controller design, it is of interest to linearize the
models with respect to the system inputs TIVC , [O2 ]IVC and θSOI , at an arbitrary
operating point
¡ 0 ¢T
X 0 = TIVC [O2 ]0IVC θSOI
0
(2.65)

to obtain a linear cycle-to-cycle model on the form

∂τ(X 0 )
τ(k + 1) = τ(k) + ∆TIVC (k)
∂TIVC
∂τ(X 0 ) ∂τ(X 0 )
+ ∆[O2 ]IVC (k) + ∆θSOI (k) (2.66)
∂[O2 ]IVC ∂θSOI

where ∆ is the forward-difference operator and k denotes cycle index. The li-
near model could then easily be incorporated in controller design frameworks

61
Chapter 2. Modeling

Model-Prediction Performance L2S1

1.5
τVAL
τVAL
1.4
M1
M2
τ [ms] Cross-Validation Data

1.3

1.2

1.1

0.9
0 1,000 2,000 3,000 4,000 5,000 6,000
cycle [-]

Figure 2.12 Modeled τ together with τVAL from operating point L2S1. The model
mismatch at cycle 1400 is an example of where θSOI was increased and the models
underestimated τVAL . The models were overall better at predicting τVAL during
variations in TIVC , [O2 ]IVC and θSOI for early θSOI .

such as linear quadratic Gaussian (LQG) control [Kalman, 1960] or linear model
predictive control [Maciejowski, 2002]. Linearization of M1, M2 and M3 can be
conducted by approximating the partial derivatives numerically

∂τ(X 0 ) τ(X 0 + ∆X ) − τ(X 0 )

≈ (2.67)
∂X ∆X

The computational complexity of (2.67) in terms of arithmetic operations are

presented in Table 2.2 for the different models. Addition, subtraction, multipli-
cation and taking powers were counted as one operation. In Table 2.2, N is the
number of samples between θSOI and θ10 , τ/h is the ratio of τ to the step size
h used in the forward-Euler approximation when simulating M3. From the au-
thor’s experience, h should be smaller than 0.05 ms. The values in Table 2.2 are
by no means shown to be the most efficient implementation of (2.67), neverthe-
less, they originate from the implementation used in this work. The results show
that M1 and M2 are superior in terms of simplicity, at least if (2.67) should be
computed by a controller every engine cycle.

62
 2.4 Ignition-Delay Modeling

Model-Prediction Performance L1S2

2.8
τVAL
2.6 τVAL
M1
2.4 M2
τ [ms] Cross-Validation Data

2.2

1.8

1.6

1.4

1.2
0 1,000 2,000 3,000 4,000 5,000 6,000
cycle [-]

Figure 2.13 Modeled τ together with τVAL from operating point L1S2.

Discussion
The models could not accurately detect when an increase in θSOI started to give
an increase in τ. If this was caused by errors in the estimated cylinder-gas state,
errors in the heat-release analysis or by the fact that the models did not account
for spatial and physical effects is not known. Model performance was however
more acceptable when θSOI was advanced earlier during the compression stroke
and when TIVC and [O2 ]IVC were varied. This implies that it is easier to obtain sa-
tisfactory controller performance when regulating τ using TIVC and [O2 ]IVC , un-
less θSOI is sufficiently far away from TDC. When evaluating performance for all
operating points simultaneously (see Fig. 2.16), model performance decreased
with complexity. Small improvements could be obtained when using four local
models instead of one global model.
The results were somewhat unambiguous when comparing model perfor-
mance, and it was only worthwhile to increase model complexity in some ca-
ses, (see L1S1, L2S1, Fig. 2.10). When studying Figs. 2.11-2.14, it can be seen
that model performance did not differ significantly in most cases. When taking
into account for the computational cost of linearization, M1 was superior to M2
and M3 (see Table 2.2). Furthermore, M1 had a closed-form expression (2.61) for

63
Chapter 2. Modeling

Model-Prediction Performance L2S2

4.5
4 τVAL
3.5 τVAL
M1
3
M2
2.5

1.4
τ [ms] Cross-Validation Data

1.2

0.8

0 1,000 2,000 3,000 4,000 5,000 6,000

cycle [-]

Figure 2.14 Modeled τ together with τVAL from operating point L2S2. Note that
the models were able to predict the steep increase in τ as θSOI was decreased
around cycle 5250. The reason for this was that cycles 4000-6000 at L2S2 were
included τID , see Fig. 2.9.

parameter calibration. Based on these findings, M1 was the suitable choice for
model-based controller design, and was therefore used by the model predictive
controller in Chapter 7.
These models did not include effects from cylinder-wall temperature, fuel va-
porization, atomization and fuel-spray interaction with the combustion cham-
ber walls. If carefully modeled, these effects are believed to improve the model
performance. However, information from these effects were not easily accessible
from the sensors available, making the validation of these effects difficult.

M4
The ignition-delay models presented above did not take fuel reactivity into ac-
count. One of the main motivators for closed-loop combustion control is the
possibility to handle fuel variation. The octane-number dependent τ correlation

64
 2.4 Ignition-Delay Modeling

All Operating Points

0.1
without p with p without p & individual OP

0.09

0.08
RMSE [ms]

0.07

0.06

0.05
M1 M2 M3

Figure 2.15 RMSE with respect to validation data. Increasing the model comple-
xity from M1 to M2 and M3 increased RMSE, and using individual operating-point
models gave a slight reduction in RMSE. Increased prediction performance could
therefore be obtained by switching between different models instead of using one
model for all operating points.

presented by Delvescovo et al. [2016]

ζ(T,PRF) Λ(T,PRF)
τ = φα(T,PRF) p β(T,PRF) xO e (2.68)
2

was therefore used in Chapter 6 to investigate how the fuel reactivity affects con-
troller design. This model was developed for primary-reference fuel (PRF) blends
from PRF0 (n-heptane) to PRF100 (iso-octane) and was calibrated with data from
constant-volume simulations in Cantera, using a reduced gasoline-surrogate ki-
netic mechanism. Note that the PRF value and octane number are equal per de-
finition. In (2.68), φ denotes equivalence ratio, p denotes pressure, and xO 2 is the
mole fraction of oxygen of the inducted gas. The parameters α, β, ζ and Λ are
temperature-dependent polynomials multiplied with a PRF-dependent expo-
nential expression, yielding roll-off at low temperatures. Delvescovo et al. [2016]
calibrated the polynomial coefficients with respect to simulated τ data at the fol-
lowing conditions

• initial temperatures from 570-1860 K

• initial pressures from 10-100 bar

• mole fractions of oxygen from 12.6 % to 21 %

65
Chapter 2. Modeling

Model Prediction Performance for all Operating Points

4
τVAL
3.5 τVAL
M1
τ [ms] Cross-Validation Data

M2
3

2.5

1.5

0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4
cycle [-] ·104

Figure 2.16 Modeled τ together with τVAL for all operating points.

• equivalence ratios from 0.30-1.5

• PRF blends from PRF0 to PRF100

to capture a wide range of engine-like operating conditions. With (2.68), the start
of combustion θSOC was computed using a Livengood-Wu integration criterion
[Livengood and Wu, 1955]
Z θSOC 1
dt = 1 (2.69)
θSOI τ

2.5 NOx -formation Model

Oxides of nitrogen NOx are regulated engine pollutants which are mainly produ-
ced by reactions between nitrogen and oxygen during combustion and expan-
sion. A NOx model was used both as a virtual sensor to provide a controller with
the previous-cycle NOx -emission level, and as a predictive model for NOx regula-
tion below an upper limit. For these purposes, the computation time needs to be
sufficiently short for cycle-to-cycle control. A model that fulfills this requirement

66
 2.5 NOx -formation Model

is the two-zone (burned and unburned gas) combustion model coupled with the
Zeldovich mechanism
O + N2 → NO + N
(2.70)
N + O2 → NO + O
that was previously developed for combustion-control purposes in [Egnell, 2001;
Murić et al., 2014]. This model assumes NOx formation through the thermal-
formation path alone. Less dominant formation paths include reactions between
nitrogen and unburnt hydrocarbons and reactions with nitrogen components
in the fuel [Fenimore, 1971]. More detailed NOx -formation models incorporate
fuel-air mixing and the cylinder air-to-fuel ratio distribution, which affect the
combustion temperature. An extensive NOx -modeling review can be found in
[Miller and Bowman, 1989].
The model used here separates the cylinder gases into a burned zone and
an unburned zone. The unburned zone contains fuel, air and EGR, whilst the
burned zone contains combustion products. When fuel and air react at a lo-
cal air-fuel ratio, λlocal , the reaction products are moved from the unburned
zone to the burned zone where combustion chemistry is assumed to be in
equilibrium after combustion. The local air-fuel ratio λlocal was here used as a
model-calibration parameter. Under these assumptions, the mass m bz and tem-
perature Tbz of the burned zone change every crank-angle increment ∆θ accor-
ding to
1 dQ c (θ)
m bz (θ + ∆θ) = m bz (θ) + (1 + λlocal + r EGR )∆θ
Q LHV d θ
µ ¶γ−1 (2.71)
p(θ + ∆θ) γ 1 dQ c (θ)
Tbz (θ + ∆θ) = Tbz (θ) + ∆θ
p(θ) m bz (θ)c p d θ
where dQ c /d θ is the heat-release rate, and r EGR is the EGR ratio. The
unburned-zone temperature Tuz changes due to isentropic compression and
expansion alone
µ ¶γ−1
p(θ + ∆θ) γ
Tuz (θ + ∆θ) = Tuz (θ) (2.72)
p(θ)
The NO-formation rate in the burned zone is given by
³ ³ [NO] ´2 ´
2r 1 1 −
d [NO] [NO]e [NO] dV
= − (2.73)
dt [NO] r 1 V dt
1+
[NO]e r 2
according to the Zeldovich mechanism [Zeldovich et al., 1947]. This expression
assumes radical species in equilibrium where NO and O equilibrium concentra-
tions, [NO]e and [O]e , can be computed from combustion-product composition,

67
Chapter 2. Modeling

temperature and equilibrium concentrations of the water-gas shift reaction. The

last term in (2.73) is due to the time-varying cylinder volume. and the reaction
rates r 1 and r 2 are given by

r 1 = 7.6 × 1013 [O]e [N2 ]e e −38000/Tbz

(2.74)
r 2 = 1.5 × 109 [NO]e [O]e e −19500/Tbz

Formation of NO2 is typically an order of magnitude smaller than NO and was

therefore not modeled explicitly. The model could however anticipate the total
amount of NOx since NO2 is formed from NO. For a more detailed description of
this model, see [Egnell, 2001; Murić et al., 2014].

Model Calibration
The NOx -model was calibrated with respect to experimental data. Figure 2.17
shows modeled NOx coincidence with measured NOx concentrations, obtained
from engine experiments. The data originate from operating points at p IMEP =
5 and 10 bar with 0 and 30 % EGR, with combustion timings from 0 to 15 CAD,
pilot-injection durations from 0 to 0.4 ms, and with diesel (×) and gasoline (◦)
fuel.
In Fig. 2.17, performance was found to degrade for low NOx concentrations,
especially with high EGR ratios where the model underestimated the measured
NOx emissions. An explanation for this could be that other formation paths be-
come more important at these operating points. Possible improvement could be
obtained by also including the N2 O formation path as described in [Gong and
Rutland, 2013], which improves NOx prediction at lower combustion tempera-
tures. It can also be seen that model errors were larger for gasoline compared to
diesel. The reason for this is unknown and deserves further investigation.

68
 2.5 NOx -formation Model

NOx -model fit, Average Error = 22.5052 %

2,500
Measured NOx [ppm]
2,000

1,500

1,000

500

500 1,000 1,500 2,000 2,500

Modeled NOx [ppm]

100
, × EGR = 0 %, Load = 5 bar
, × EGR = 0 %, Load = 10 bar
75
, × EGR = 30 %, Load = 5 bar
, × EGR = 30 %, Load = 10 bar
Error [%]

50

25

0
500 1,000 1,500 2,000 2,500
Measured NOx [ppm]

Figure 2.17 The NOx model was calibrated at p IMEP = 5 and 10 bar and 1200
rpm, with 0 and 30 % EGR, combustion timings from 0 to 15 CAD, pilot-injection
durations from 0 to 0.4 ms, and with diesel (×) and gasoline (◦) fuel.

69
3
Control and Estimation
Methods

This chapter gives a brief overview of the control and state-estimation methods
used in the thesis.

3.1 Control Concepts

PI Control
The PI (Proportional Integral) controller is the most common solution to practi-
cal control problems [Åström and Hägglund, 2006]. A discrete-time PI controller,
suitable for cycle-to-cycle control, is given by
k−1
X
u(k + 1) = k p e(k) + k I e(i ) (3.1)
i =0

where u is the controller output, e is the control-error, i.e., the system-output

deviation from a set point, k is the cycle index, and k p and k I are controller gains.
The controller in (3.1) is obtained by discretizing the continuous PI controller
counterpart with a forward-Euler approximation.
The PI controller was used for many different purposes in this work. It was
used to control p IMEP by adjusting the fuel-injection duration, and to control the
common-rail pressure p rail by adjusting the inlet-metering valve position, which
determines the fuel-flow to the common-rail volume. Chapter 6 investigates ro-
bust PI-controller tuning for combustion-timing control, and a hybrid PI control-
ler was used in Chapter 10 for simultaneous tracking and constraint fulfillment
with multiple injections.
PI controllers were used because of their simplicity with respect to tuning and
implementation. A simple controller design is attractive if performance require-
ments are met. Previous research have also shown that PI control is a suitable
option for combustion-control purposes [Bengtsson, 2004; Hanson, 2010].

70
 3.1 Control Concepts

Model Predictive Control

Model predictive control (MPC) is a control technique where a system model

x(k + 1) = f (x(k), u(k))

(3.2)
y(k) = h(x(k))

is used to solve a finite horizon optimal-control problem every sample. Here, we

denote the system state x ∈ Rn , the system input u ∈ Rm and the measured output
y ∈ Rp . The solution to the optimal-control problem is an open-loop sequence
U = {u(k)} from k = 1 to Hc , where Hc is the control horizon, and k = 0 is the
current sample. The first input of the optimal open-loop sequence u ∗ (1) is then
repeatedly actuated. The input sequence is optimal in the sense that it minimizes
a cost function
Hp Hc
X X
J (u(k), y(k)) = J y (y(k)) + J u (u(k)) + ρ ² ²2 (3.3)
k=1 k=1

which consists of output costs J y (k) over a prediction horizon H p and input costs
J u (k) over a control horizon Hu . The output cost at sample k
p
X
J y (y(k)) = ω y j (y rj (k) − y j (k))2 (3.4)
j =1

is the sum of the squared set-point deviations y rj (k) − y j (k) for each output y j
and corresponding set point y rj , scaled with a cost weight ω y j . The input cost at
sample k
Xm
J u (u(k)) = ωu j u j (k)2 + ω∆u j ∆u j (k)2 (3.5)
j =1

is the square sum of both absolute input values u j and changes ∆u j with corres-
ponding weights ωu j and ω∆u j . The positive cost weights ωx are controller design
parameters that determine how the controller prioritizes different output errors
and control actions.
Feedback is introduced by minimizing (3.3) subject to the system initial con-
ditions x(0) = x 0 , obtained from measurement or state estimation. The main fea-
ture of MPC is its ability to incorporate constraints with respect to inputs, outputs
and states in the optimization problem
y y
−²η min + y min ≤ y(k) ≤ y max + ²η max
−²ηxmin + x min ≤ x(k) ≤ x max + ²ηxmax
(3.6)
−²ηumin + u min ≤ u(k) ≤ u max + ²ηumax

−²η∆u ∆u
min + ∆u min ≤ ∆u(k) ≤ ∆u max + ²η max

71
Chapter 3. Control and Estimation Methods

yr

k =0 k = Hp

Figure 3.1 Illustration of the model predictive control principle for a

one-dimensional system, where H p = Hc . At the current sample k = 0 with me-
asured output y, the computed future optimal control sequence u gives an op-
timal output sequence with respect to the prediction horizon H p (dash-dotted).
The first output of the optimal input sequence is then actuated to the system at
the next time step. This procedure is then repeated every sample. In this example,
the objective is to track a set-point signal y r subject to constraints on the input
(dashed).

The positive variable ² and vectors η determine costs for constraint violation.
When η = 0, the solution U ∗ is not allowed to violate the constraint, but when
η 6= 0, the constraint can be violated with additional cost, determined by the cost
weight ρ ² in (3.3). The design choice of having η > 0 was used in Chapter 8 to en-
sure existence of feasible solutions. The principle of MPC is illustrated in Fig. 3.1
for a single-input / single-output system where the objective is to track a set point
y r , subject to constraints on the input (dashed).
If the system dynamics are linear, the optimization problem of minimizing
(3.3) subject to the equality constraints in (3.2), and the inequality constraint in
(3.6) is a quadratic-programming (QP) problem
1 T
minimize x Hx + f T x (3.7)
x 2
subject to Ax ≤ b
A eq x = b eq

A QP is a convex optimization problem, which means that it has many benefi-

cial properties [Boyd and Vandenberghe, 2008]. The most important being that
a local optimal point is also globally optimal. A QP can be solved efficiently with
custom-made solvers, often within a few milliseconds and sometimes even fas-
ter [Mattingley and Boyd, 2012]. In this thesis, QPs were obtained by linearizing
the system equations, see below. Linearization was then recomputed every en-

72
 3.1 Control Concepts

gine cycle to update the optimization-problem approximation. More details on

the MPC implementation are presented in Chapter 5.
MPC has previously been applied to a wide range of problems including
chemical-process control [Maciejowski, 2002], supply-chain management [Cho
et al., 2003] and finance [Dawid, 2005]. MPC has also been used to con-
trol HCCI and PPC engines where it was chosen primarily for its MIMO and
constraint-handling capabilities [Bengtsson et al., 2006; Widd et al., 2009; Lewan-
der et al., 2008].
In this work, MPC was chosen for its ability to handle input and output con-
straints. For the control problems investigated, there were constraints on valve
positions, cylinder pressure, exhaust temperature and NOx emissions. With ac-
tive constraints, MPC obtains nonlinear properties that are not obtainable with
standard linear controller designs such as LQR or PID [Maciejowski, 2002]. The
MPC framework is also convenient when the system is large, as in Chapter 7. In
this case, tuning of the cost function weights ωx is an intuitive way of formula-
ting the controller objectives. An optimization framework is also well suited for
engine control, since the controller objective does not always involve set-point
tracking. An example of this is investigated in Chapter 11, where the objective is
to maximize η GIE subject to pressure constraints.

Linearization
Linear model predictive control requires linear system models. Most of the mo-
dels presented in Chapter 2 are however nonlinear. In order to enable linear con-
troller design, one can linearize the system dynamics at the current operating
point. Given a nonlinear discrete-time model on the form

x(k + 1) = f (x(k), u(k))

(3.8)
y(k) = h(x(k))

a linear model approximation

∆x(k + 1) = A∆x(k) + B ∆u(k)

(3.9)
∆y(k) = C ∆x(k)

is obtained from the partial derivatives

∂ fi ∂ fi
Ai j = (x 0 , u 0 ), Bi j = (x 0 , u 0 )
∂x j ∂u j
(3.10)
∂h i
Ci j = (x 0 , u 0 )
∂x j

where ∆x(k), ∆u(k) and ∆y(k) are deviations from the linearization point x 0 , u 0 ,
y 0 . Linearization was used in Chapter 7 to linearize an ignition-delay model, and

73
Chapter 3. Control and Estimation Methods

in Chapter 11 to linearize the cylinder-pressure model presented in Chapter 2.

Since the linear approximation is only accurate close to the linearization point,
(3.10) was recomputed every engine cycle.

3.2 State Estimation

The objective in state estimation is to estimate the system state x given measu-
rements y. With the probabilistic representation of a dynamic system with state
x(k), measured output y(k) and input u(k)

x(k + 1) ∼ p(x(k + 1)|x(k), u(k))

y(k) ∼ p(y(k)|x(k)) (3.11)
x 0 ∼ p(x 0 )

the state-estimation problem amounts to evaluate the probability density func-

tion of x(k) given previous measurement y(k)

p(x(k)|y(k), y(k − 1), . . . , y(0)) (3.12)

Here, the notation x ∼ p(x|y) indicates that x is distributed according to the con-
ditional probability-density function p of x, given y.

The Kalman Filter

For the case when the system in (3.11) is linear

x(k + 1) = Ax(k) + B u(k) + v (k)

(3.13)
y(k) = C x(k) + e(k)

and perturbed with additive Gaussian noise with zero mean and covariance ma-
trices Q and R
v (k) ∼ N (0,Q)
(3.14)
e(k) ∼ N (0, R)

the estimation problem has an analytic solution, and the Kalman Filter (KF)
provides an optimal algorithm for iteratively estimating x(k), see Algorithm 1
[Kalman, 1960]. The Kalman filter is used in Chapter 7 to estimate the EGR mass
flow for [O2 ]IVC computation, and in Chapter 8 to filter cycle-to-cycle variation in
pressure-rise rate and combustion timing.
If the system (3.11) is nonlinear on the form

x(k + 1) = f (x(k), u(k)) + v (k)

(3.15)
y(k) = h(x(k)) + e(k)

74
 3.2 State Estimation

Algorithm 1 Kalman Filter

1: Initialize x̂(0) and P (0)

2: while k > 0 do
3: x̂(k|k − 1) = A x̂(k − 1|k − 1) + B u(k − 1)
4: P (k|k − 1) = AP (k − 1|k − 1)A T +Q
5: e(k) = y(k) −C x̂(k|k − 1)
6: S(k) = C P (k|k − 1)C T + R
7: K (k) = P (k|k − 1)C T S −1 (k)
8: x̂(k|k) = x̂(k|k − 1) + K (k)e(k)
9: P (k|k) = (I − K (k)C )P (k|k − 1)
10: end while

the Kalman filter is unfortunately not directly applicable. There are however
well-established methods for solving nonlinear estimation problems. The exten-
ded Kalman filter (EKF) and the particle filter (PF) are examples of such methods.
They were used in Chapter 4 for automatic calibration of the heat-release model
presented in Chapter 2.

The Extended Kalman Filter

The EKF provides an approximate solution to the estimation problem by linea-
rizing (3.15) at the current estimate in order for the iterative KF procedure to be
applied [Julier and Uhlmann, 2004]. The EKF is presented in Algorithm 2.

Algorithm 2 Extended Kalman Filter

1: Initialize x̂(0) and P (0)

2: while k > 0 do
3: x̂(k|k − 1) = f (x̂(k − 1|k − 1)|u(k − 1))
∂f
4: A(k − 1) = ∂x |x̂(k|k−1),u(k−1)
5: P (k|k − 1) = A(k − 1)P (k − 1|k − 1)A(k − 1)T +Q
6: e(k) = y(k) − h(x̂(k|k − 1))
∂h
7: C (k) = ∂x |x̂(k|k−1)
8: S(k) = C (k)P (k|k − 1)C T (k) + R
9: K (k) = P (k|k − 1)C (k)T S −1 (k)
10: x̂(k|k) = x̂(k|k − 1) + K (k)e(k)
11: P (k|k) = (I − K (k)C (k))P (k|k − 1)
12: end while

75
Chapter 3. Control and Estimation Methods

The Particle Filter

The PF is a sequential Monte Carlo sampling method that aims to approximate
the conditional distribution in (3.12) numerically. The PF consists of a set of N p
sampled particles x i (k) with corresponding weights ωi . Together, they provide a
point-mass approximation

Np
X
p(x(k)|y(k)) ≈ ωi (k)δ(x(k) − x i (k)) (3.16)
i =1

where δ(·) is the Dirac delta function. The PF performs the sampling proce-
dure using a sequential Monte Carlo technique where particles x i (k + 1) are
sampled sequentially, given the old particles x i (k), and a proposal distribution
q(x(k + 1)|x i (k)). After each time step, the weights are updated to represent the
desired probability density function in (3.12). When choosing q(x(k + 1)|x i (k)) =
p(x(k + 1)|x i (k), u(k)), the update rule for the weights becomes ωi (k + 1) =
ωi (k)p(y(k + 1)|x i (k + 1)). To avoid particle depletion, meaning that only a few
weights contribute to (3.16), the particles have to be resampled according to the
weight distribution. This procedure puts more particles into areas of high proba-
bility and discards particles in regions of low probability. The resampling step is
commonly conducted when the ratio of the number of effective particles Neff

1
Neff = PN (3.17)
p
i =1
(ωi )2

to N p becomes too low. This estimation method is referred to as the bootstrap PF

and was first introduced by [Gordon et al., 1993], see Algorithm 3.

Algorithm 3 The Particle Filter

1: Draw N p particles x i (0) from the initial distribution p(x(0)) and initialize the
weights ω(0)i = 1/N p
2: while k > 0 do
3: Update the particles x i (k) by sampling p(x(k)|x i (k − 1), u(k − 1))
4: Update the weights ωi (k) = ωi (k − 1)p(y(k)|x i (k)) and renormalize
5: if Neff < x frac N p then
6: Draw new particles from the distribution defined by {ωi (k)}i =1...Np
7: Reinitialize the weights ωi (k) = 1/N p
8: end if
9: end while

76
4
Heat-Release Analysis

4.1 Introduction
Heat-release analysis refers to the use of physical models, such as the pres-
sure model introduced in Chapter 2, to infer information about the combustion
processes from cylinder-pressure measurements. This information is commonly
used for engine diagnostics, research, control and simulation. Pioneering work
on methods for inferring the heat-release rate from measured in-cylinder pres-
sure were presented in [Rassweiler and Withrow, 1938; Krieger and Borman, 1966;
Gatowski et al., 1984]. From a feedback-control perspective, heat-release analy-
sis provides the possibility to regulate combustion timing and ignition delay on
a cycle-to-cycle basis.

Heat-Release Analysis
In this work, the heat-release rate dQ c /d θ is computed from the measured pres-
sure p by rearranging (2.9)

dQ c γ dV 1 d p dQ ht
= p + V + (4.1)
dθ γ − 1 dθ γ − 1 dθ dθ

Crank angles of x% burned θx are commonly used indicators for the timing and
duration of the combustion process. These are obtained from the accumulated
heat release Z θ
dQ c
Q c (θ) = dθ (4.2)
θIVC d θ

and the relation

Q c (θx )
x = 100 (4.3)
maxQ c (θ)
θ
In this work, θ10 was used to compute τ and θ50 was used to indicate combustion
timing. The difference θ90 − θ10 is a commonly used measure for the duration of
combustion. Figure 4.1 illustrates how these quantities are obtained from mea-
sured p and estimated dQ c /d θ.

77
Chapter 4. Heat-Release Analysis

6,000 p

Qc
4,000
Q c [J]

2,000
θ50
0

θ SOI τ θ10 θ90

−2,000
−20 −10 0 10 20 30 40
θ [CAD]

Figure 4.1 The accumulated heat-release Q c is obtained from the measured in

cylinder pressure p and (4.1). It provides important feedback variables, such as
the combustion timing θ50 and the ignition delay τ.

With multiple large injections and a multimodal heat-release rate, it is no lon-

ger sufficient to have θ50 as a combustion-timing indicator. Peak detection was
i
therefore used instead to extract combustion timings θCT , under the assumption
that each injection contributes to a heat-release impulse. Here, i denotes injec-
tion index where i = 1, . . . , M , with M injections. Detection was conducted by
computing the M largest maxima of dQ c /d θ, larger than a threshold dQ t , and
with a minimum separation in θ. The separation criterion was included since dif-
fusion combustion has a characteristic double-peak heat-release rate that could
result in additional peaks detected. The peak-detection method also accounted
for the ordering of injections when allocating combustion timings to injections.
The use of threshold criteria for detection is a commonly used approach in sig-
nal processing, where the threshold-magnitude used is a trade-off between the
probabilities of having false positives and false negatives [Kay, 1998].
It is sometimes necessary to separate dQ c /d θ among different injections. For
this purpose, dQ c /d θ was assumed to consist of the heat-release generated from
the different injections according to

dQ c X M dQ i
c
= (4.4)
dθ i =1 d θ

The procedure used for obtaining dQ ci /d θ from dQ c /d θ was to first detect the
M most significant peaks, as described above. With the peaks detected, dQ c /d θ

78
 4.1 Introduction

was separated in different intervals according to


dQ c
d Q̂ ci  if l i ≤ θ ≤ d i
= dθ (4.5)
dθ 
0 otherwise

where the bounds l i and d i were determined from computed minima

in-between detected peaks. In order to obtain more physical heat-release shapes,
d Q̂ i /d θ were smoothed through a zero-phase filter, and normalized so that (4.4)
is fulfilled
dQ ci ³ X
M d Q̂ i ´−1 d Q̂ i dQ
c c c
= (4.6)
dθ i =1 d θ d θ d θ
The separation method is illustrated in Fig. 4.2 for a multimodal heat-release rate
with three injections. This method is later used in chapters 10 and 12, where com-
bustion feedback with multiple injections is investigated.
The proposed combustion-detection method is fairly simplistic and was
developed out of necessity. Suggested further development is to instead use a
statistical method where a probabilistic combustion model is utilized in combi-
nation with fuel-injection information.

Model Parameters
The heat-release model in (4.1) has a set of unknown parameters that have to
be tuned for satisfactory performance. These parameters can be tuned man-
ually with knowledge of the appearance of physical heat-release rates and the
influence of the different model components. This procedure is, however, time
consuming and has to be redone from time to time.
The development of automatic calibration methods has been an ac-
tive research area during the past decades, where methods for calibrating
pressure-sensor offset [Tunestål et al., 2001; Brunt and Pond, 1997], polytro-
pic coefficients [Manente et al., 2008; Randolph, 1990], volume-curve offset
[Stas, 2004; Tunestål, 2001] and compression ratio [Klein et al., 2006] have been
presented. In [Klein, 2007], an off-line method for calibration of a large set of
parameters simultaneously was presented and studied in detail. It was, however,
concluded by Eriksson [1998], that all model parameters might not be identifia-
ble simultaneously. This indicates that a calibration problem involving a large
set of parameters is not easily solved.
This chapter investigates on-line calibration of a subset of the model parame-
ters in (4.1). The task is first formulated as a nonlinear estimation problem, where
unknown states of a dynamic system are to be estimated given a statistical mo-
del and sensor measurements. The formulated estimation problem is then sol-
ved using the extended Kalman filter (EKF) and the bootstrap particle filter (PF),

79
Chapter 4. Heat-Release Analysis

600 600
dQ c /d θ [J/CAD] 1. 2.

dQ c /d θ [J/CAD]
400 400

200 200

0 0
−20 0 20 −20 0 20
θ [CAD] θ [CAD]

600 600

3. 4.
dQ c /d θ [J/CAD]

dQ c /d θ [J/CAD]
400 400

200 200

0 0
−20 0 20 −20 0 20
θ [CAD] θ [CAD]

Figure 4.2 A method for separating dQ c /d θ among several injections. The

heat-release rate is first obtained from the measured pressure signal (1). The M
most significant peaks are then detected (2). The detected peaks constitute the
combustion timings θCT i . The heat release is then separated in different intervals

(3) according to detected peak locations. The heat-release rates d Q̂ ci /d θ in the

different intervals are then filtered and normalized (4).

described in Chapter 3. These filters lend themselves nicely for real-time appli-
cations because of their sequential processing of measured data. The problem
representation provides a general framework for which any parameter combina-
tion could be estimated as long as the system is observable.
The chapter is outlined as follows: The estimation-problem formulation and
filter configurations are introduced in Secs. 4.2 to 4.3. Filter-performance results
with respect to simulated and experimental data are then given in Secs. 4.4 to 4.5.
Discussion and conclusions are presented in Secs. 4.6 and 4.7.

80
 4.2 Problem Formulation

4.2 Problem Formulation

The problem considered was to estimate the TDC offset θ∆TDC , the convective
heat-transfer coefficient C 2 (see (2.16)), and the cylinder-wall surface tempera-
ture T w from cylinder-pressure data. The heat-release model is given by

dQ c γ dV 1 d p dQ ht
= p + V + (4.7)
dθ γ − 1 dθ γ − 1 dθ dθ

where γ is determined from NASA polynomials and dQ ht /d θ is given by the Wos-

chni heat-transfer model in (2.14-2.16). It was concluded in [Brunt, 1997] that
the p IMEP error from θ∆TDC is 3 to 10 % per CAD. The parameter C 2 was chosen
over C 1 due to its greater impact on heat-transfer rate as shown in [Klein, 2008].
The wall-surface temperature T w , has previously been shown to be an important
state for describing the slower combustion-timing dynamics in low temperature
combustion [Blom 2008]. If these parameters are set correctly, the accumulated
heat release should be zero before the start of combustion θSOC and constant at
the value of released fuel energy after the end of combustion θEOC . The validity
of this assumption is affected by measurement noise and model errors. Given the
correct model-parameters, the computed accumulated heat release is assumed
to be on the form
½
²1 θ ≤ θSOC
Q c (θ) = (4.8)
Q ctot + d + ²2 θ ≥ θEOC

where ²1 ∼ N (0, σ21 ) and ²2 ∼ N (0, σ22 ) are i.i.d normally distributed noise proces-
ses, with standard deviations σi . They are introduced to represent sensor noise
and unmodeled effects. The mean injected fuel energy Q ctot is assumed to be
known and determined by the injection duration and common-rail pressure. The
variable d ∼ N (0, σ2d ) is a random offset accounting for stochastic variation in the
injected fuel energy.

Nonlinear State-Space Model with Gaussian Noise

Prior knowledge of the unknown states was modeled as a Gaussian distribution

¡ 0 0
¢T
C2 θ∆TDC T w0 ∼ N (µ0 , Σ0 ) (4.9)

where µ0 is an initial guess with corresponding covariance Σ0 . Dynamic

cycle-to-cycle variation was represented by the state-space model
      
C 2 (k + 1) 1 0 0 C 2 (k) 0
   0    
θ∆TDC (k + 1) = 0 1  θ∆TDC (k) +  0  + v (k) (4.10)
T w (k + 1) 0 0 Φ T w (k) Γ1 T + Γ2 Tc

81
Chapter 4. Heat-Release Analysis

6,000
Qc θ EOC
Q ctot
4,000
Q c [J]

2,000
θ SOC

−60 −40 −20 0 20 40 60 80

θ [CAD]

Figure 4.3 A realization of the statistical model in (4.9-4.13). The dashed line re-
presents the expected accumulated heat release before and after combustion wi-
thout stochastic variation. Deviation from this value is due to an offset d and noise
processes ²1 , and ²2 for which σ21 = 0, σ22 = 2500. It is assumed that the injected
fuel energy Q ctot , the start and end of combustion θSOC and θEOC are available.

where T w evolves according to the heat-transfer model presented in Sec. 2.2

H
Φ = e A w (θ)d θ
I Rϑ (4.11)
A (ϑ)d ϑ
Γ1 T + Γ2 Tc = e θBDC w B w (θ)(T + Tc )d θ

The uncertainty in C 2 and θ∆TDC was modeled as a random walk driven by a

Gaussian process

v (k) ∼ N (0, Σ1 ) (4.12)

The covariance matrix in (4.12), Σ1 , is a measure of the model uncertainty, and

was here chosen as a diagonal matrix with elements representing the state un-
certainty. The model generates a Q c output every engine cycle according to

dQ c γ dV 1 d p dQ ht
= p + V +
dθ γ − 1 dθ γ − 1 dθ dθ
Z (4.13)
θ dQ c
Q c (θ) = dθ
θIVC dθ

82
 4.2 Problem Formulation

where a realization of (4.13) with the correct model parameters is shown in Fig.
4.3. Equations (4.9)-(4.13) are on the form

x(k + 1) = f (x(k), u(k)) + v (k)

y(k) = h(x(k)) + e(k) (4.14)
x(0) = µ0 + e 0

where the state-update equation is given by (4.10), and the system-output

equation by (4.13). The objective is now to estimate x(k) given measurements
y(k), . . . , y(0). In our case, y(k) is not obtained from sensor measurement. It is in-
stead given by the deterministic part of (4.8), which is the expected heat-release
rate before θSOC and after θEOC , given the fuel energy Q ctot .

Observability
It is crucial to evaluate the system observability when conducting state-estimation.
Observability is a measure of how well the system state x(k) can be inferred
from knowledge of the system output y(k), and was introduced to linear-system
theory by Kalman [1959]. In order to apply the concept of observability to the
system in (4.10-4.13), a linearization

∆x(k + 1) = A(k)∆x(k) + B (k)∆u(k)

(4.15)
∆y(k) = C (k)∆x(k)

was conducted at the operating point presented in Table 4.1. In our case, observ-
ability was investigated by evaluating

∂h i (x(k))
C (k)i j = (4.16)
∂x j

Since A(k) ≈ I 3x3 , and the observability matrix O is given by

 
C
O = C  (4.17)
C

The system (4.15) is observable if O , or in this case C , has 3 independent rows.

A rescaled C (k) is presented in Fig. 4.4, where it can be seen that θ∆TDC has a
relatively large asymmetric affect on Q c around TDC. Changes in C 2 and T w de-
termine the final Q c value, whereas T w also affects Q c during the compression
stroke. The linear-system approximation (4.15) is observable since the curves in
Fig. 4.4 are linearly independent. Despite this fact, it can be seen that negatively
correlated variation in C 2 and T w might be difficult to detect due to their similar
effect on Q c after TDC. Johansson [2006] presented similar results for guidance
in manual tuning of heat-release model parameters.

83
Chapter 4. Heat-Release Analysis

System Observability
150
∂h/∂θ∆TDC
100 ∂h/∂C 2 × 0.001
∂h/∂T w × 100

50

0
C

−50

−100

−150

−200
−150 −100 −50 0 50 100 150
θ [CAD]

Figure 4.4 Computed C when linearizing (4.10-4.13) at the operating point in

Table 4.1. The parameters affect Q c differently: θ∆TDC affects Q c asymmetrically
around TDC; C 2 determines the final Q c level; T w reduces Q c both before and
after TDC.

4.3 Filter Configurations

The remainder of this chapter covers implementation and evaluation of an exten-
ded Kalman filter (EKF) (see Algorithm 2) and a particle filter (see Algorithm 3)
for the purpose of estimating the unknown parameters. Partial derivatives (∂ f /∂x
and ∂h/∂x) necessary for the EKF were obtained through numerical differentia-
tion of (4.10) and (4.13). The probability density functions p(x(k + 1)|x(k), u(k))
and p(y(k +1)|x(k +1)), necessary for the PF, were given by (4.8), (4.10) and (4.13)
directly. Measurements y(k) were given by the deterministic part of (4.8).
In order to reduce computational effort, (4.8) was downsampled from a reso-
lution of 0.2 CAD to 1 CAD and considered 50 samples before θSOC and 50 sam-
ples after θEOC . Computation times below 1 ms per iteration were obtained for
the EKF with compiled Matlab code. To obtain comparable runtimes with the PF,
the number of particles N p was set to 250, which enabled the PF to run under 5
ms. The particle number N p is a trade-off between performance and computa-
tion time. This trade-off was however not investigated here. Particle resampling
was done when the number of effective particles Neff was below 0.25N p .

84
 4.4 Simulation Results

4.4 Simulation Results

The filters were evaluated with respect to consistency and rate of convergence,
sensitivity to statistical assumptions and model-parameter errors. This was done
with respect to simulated pressure traces generated from the model
µ ¶
dp γ dV γ − 1 dQ c dQ ht
=− p+ − , p(θIVC ) = p in (4.18)
dθ V dθ V dθ dθ
using the MATLAB ode23s solver, the model parameters in Table 4.1 and the noise
densities
σ1 = 25, σ2 = 25, σd = 100
 
9/4 0 0
Σ0 =  0 2.5 × 10−7 0 
0 0 625 (4.19)
 
0.0625 0 0
Σ1 =  0 1 × 10−8 0 
0 0 6.25

Convergence
Filter consistency and rate of convergence were evaluated by initializing the EKF
and the PF with correct model parameters according to Table 4.1, apart from the
incorrect initial filter states
¡ ¢T
x 10 = 1, 0.0095, 595
¡ ¢T (4.20)
x 20 = −1, 0.002, 336

with the true state being

¡ ¢T
x ∗ = 0, 0.0032, 465 (4.21)
Figure 4.5 shows simulation results where the mean filter state and standard
deviation of 25 realizations are presented as a function of engine cycle for the
initial conditions x 10 (dashed) and x 20 (solid). Note that the cycle-axis scales are
different for the two filters. It can be seen that the filters converged to the correct
state and that the EKF had a higher convergence rate, where the PF convergence
rate probably could have been improved by increasing N p . It can also be seen
that the θ∆TDC estimates had very fast initial transients, which indicates a rela-
tively high filter sensitivity to θ∆TDC errors. Filter RMSE at cycle k
v
u N
u1 X
RMSE(k) = t (x̂ i (k) − x ∗ )2 (4.22)
N i =1

is indicated by the black lines in Fig. 4.6 for N = 25.

85
Chapter 4. Heat-Release Analysis

Table 4.1 Nominal operating point for filter evaluation. The heat-release rate,
dQ c /d θ, was chosen as a Gaussian function with a standard deviation of 5 CAD.

Boundary Conditions Combustion Properties

p in [bar] 1.3 θ50 [CAD] 5
dQ c /d θ
Tin [K] 303 [-] N (5, 5)
Q ctot
λ [-] 2 Q ctot [J] 4 × 103
r EGR [-] 0
Tex [K] 400
Cylinder Geometry Heat-Transfer Parameters
r c [-] 16 C 1 [-] 2.28
3 3
Vd [m ] 2.1 × 10 C 2 [m/(sK)] 0.0032
B [mm] 130 T w [K] 465
L [mm] 160 Tc [K] 333
IVC [CAD] -151 m c c p [J/K] 1150
EVC [CAD] 146 k c [J/(mK)] 45
θ∆TDC [CAD] 0 L c [m] 0.025

Sensitivity to Model Uncertainty, Σ1

Sensitivity to the assumed model uncertainty was investigated by scaling the
covariance matrix of the noise term acting on (4.10) Σ1 a factor of 16. The resul-
ting RMSE (red) is compared to the nominal RMSE (black) in Fig. 4.6. An increase
in the assumed model uncertainty gave a higher convergence rate, but a slightly
larger steady-state error.

Sensitivity to Q c Noise Levels, σx

Sensitivity to the assumed Q c noise variance was investigated by scaling σ1 , σ2
and σd a factor of 4. The resulting RMSE is indicated by the red lines in Fig. 4.7 to-
gether with the nominal RMSE in black. An increased assumed Q c noise variance
clearly decreased the convergence rate.
p Sensitivity to σd was also tested by setting σd = 0 whilst σ2 was set to
252 + 1002 . This corresponds to precise knowledge of the injected fuel energy
but with increased variation from unstructured, Gaussian uncertainty. The resul-
ting RMSE is shown in Fig 4.8, where convergence with the modified noise model
is shown in red. Assuming Gaussian noise before and after combustion clearly

86
 4.4 Simulation Results

EKF PF
1 1
θ ∆ TDC [CAD]

θ ∆ TDC [CAD]
0.5 0.5

0 0

−0.5 −0.5

−1 −1
0 20 40 60 80 0 50 100 150 200
cycle [-] cycle [-]

·10−2 ·10−2
1 1
0.8 0.8
C 2 [m/(sK)]

C 2 [m/(sK)]
0.6 0.6
0.4 0.4
0.2 0.2
0 0
20 40 60 80 50 100 150 200
cycle [-] cycle [-]

600 600
T w [K]

T w [K]

400 400

200 200
20 40 60 80 50 100 150 200
cycle [-] cycle [-]

Figure 4.5 Filter convergence with initial conditions x 01 (dashed) and x 02 (solid).
True state values are indicated in red. Note the different scales on the cycle-axes
for the two filters. It can be seen that the filters converged to the correct state
(red, dash-dotted) and that the EKF had a higher convergence rate. Also note that
θ∆TDC estimates had faster initial transients. This indicates a relatively high filter
sensitivity to θ∆TDC errors.

degrades performance when there is a stochastic offset d . Another observation is

that the EKF had a large transient and a larger steady-state RMSE.
Overall, the filters behaved as one would expect. An increase in the assumed
model uncertainty increased convergence rates. For the EKF, the measurement
equation contributed more to the state estimate in relation to the model dyna-
mics, and for the PF, a larger particle spread was made each sample which incre-

87
Chapter 4. Heat-Release Analysis

EKF PF

θ ∆ TDC RMSE [CAD]

θ ∆ TDC RMSE [CAD]

0.4 0.4

0.2 0.2

0 0
0 20 40 60 80 100 0 50 100 150 200
cycle [-] cycle [-]

·10−3 ·10−3
8 8
C 2 RMSE [m/(s K)]

C 2 RMSE [m/(s K)]

6 6

4 4

2 2

0 0
0 20 40 60 80 100 0 50 100 150 200
cycle [-] cycle [-]

200 200
T w RMSE [K]

T w RMSE [K]

150 150

100 100

50 50

0 0
0 20 40 60 80 100 0 50 100 150 200
cycle [-] cycle [-]

Figure 4.6 Filter RMSE with the true model uncertainty Σ1 (black) and an incre-
ased assumed model uncertainty 16Σ1 (red). The latter increased the filter con-
vergence rate and steady state RMSE.

ased exploration of the parameter space. This did, however, increase steady-state
variation. Filter convergence rates decreased when the assumed Q c noise levels
were increased. This was due to the reduced parameter correction for Q c devia-
tion from (2.9), since variation in Q c was more probable. Assumed noise varian-
ces essentially functioned as tuning parameters that determined the trade-off be-
tween speed of convergence and steady-state variation.
Incorporating the stochastic fuel-energy component d in the filters was
shown to be important. This made constant Q c after combustion more proba-
ble and allowed for constant offsets in Q c after θEOC , which improved robustness

88
 4.4 Simulation Results

EKF PF

θ ∆ TDC RMSE [CAD]

θ ∆ TDC RMSE [CAD]

0.4 0.4

0.2 0.2

0 0
0 20 40 60 80 100 0 50 100 150 200
cycle [-] cycle [-]

·10−3 ·10−3
8 8
C 2 RMSE [m/(s K)]

C 2 RMSE [m/(s K)]

6 6

4 4

2 2

0 0
0 20 40 60 80 100 0 50 100 150 200
cycle [-] cycle [-]

200 200
T w RMSE [K]

T w RMSE [K]

150 150

100 100

50 50

0 0
0 20 40 60 80 100 0 50 100 150 200
cycle [-] cycle [-]

Figure 4.7 Filter RMSE with the correct Q c noise levels σx (black) and increased
assumed noise levels, with 4σx (red). An increased assumed Q c noise level clearly
decreased the convergence rates.

to variation in the burned fuel energy. Without d , the filters attempted to fit the
mean Q c to Q ctot after θEOC , which resulted in non-physical Q c appearances.

Sensitivity to Model-Parameter Errors

The sensitivity to erroneous model parameters was investigated by introducing
errors to p in , Tin , λ, r c , Tc and Q ctot . Stationary parameter-estimate bias due to
these errors are presented in Table 4.2, together with the model-parameter error
magnitude. It can be concluded that the filter estimates are very sensitive to er-
rors in p in and r c , but not as sensitive to errors in the other parameters. A positive

89
Chapter 4. Heat-Release Analysis

EKF PF
6 6

θ ∆ TDC RMSE [CAD]

θ ∆ TDC RMSE [CAD]

4 4

2 2

0 0
0 20 40 60 80 100 0 50 100 150 200
cycle [-] cycle [-]

·10−2 ·10−2
1.5 1.5
C 2 RMSE [m/(sK)]

C 2 RMSE [m/(sK)]
1 1

0.5 0.5

0 0
0 20 40 60 80 100 0 50 100 150 200
cycle [-] cycle [-]

400 400
T w RMSE [K]

T w RMSE [K]

300 300

200 200

100 100

0 0
0 20 40 60 80 100 0 50 100 150 200
cycle [-] cycle [-]

Figure 4.8 Filter RMSE with correct Q c noise levels

p (black) and modified assu-
med Q c noise levels (red), where σd = 0 and σ2 = 252 + 1002 . The assumption of
Gaussian noise before and after combustion clearly degraded performance when
there was a stochastic offset d to the total fuel energy. The EKF oscillated greatly
and also had larger steady-state RMSE.

error in p in or r c gives an increase in the assumed motored pressure curve p m and

a decrease in the computed T , see (2.11) and (2.16). Both effects contribute to a
decrease in dQ ht /d θ and affect estimated C 2 and T w . The parameter sensitivity
could be decreased by increasing the assumed Q c noise level σx . The numbers in
parentheses in Table 4.2 correspond to stationary parameter-estimate bias when
σx was increased a factor of 4. Another observation is that the PF was overall less
sensitive to model errors.

90
 4.5 Experimental Results

4.5 Experimental Results

The filters were also tested against experimental data with a known θ∆TDC of 1
CAD, obtained from the operating point in Table 4.3. Filter convergence can be
seen in Figs 4.9 and 4.10 when initiated with the states in (4.20), and manually
tuned noise parameters

σ1 = 100, σ2 = 100, σd = 400

 
9/4 0 0
Σ0 =  0 2.5 × 10−7 0 
0 0 625 (4.23)
 
0.0625 0 0
Σ1 =  0 1 × 10−8 0
0 0 0.1

which gave an acceptable trade-off between convergence rate and steady-state

estimate variance. The T w noise level was set low, which allowed T w to follow the
dynamic heat-transfer model whilst θ∆TDC and C 2 were adjusted for probable Q c
output.
The initial Q c output (black, dashed) is presented in Fig. 4.9, together with
the final Q c outputs (red, black, solid), where the known Q ctot level (black,
dash-dotted) was computed from fuel-flow measurements. It can be seen that
the PF converged closer to the estimated final Q c value in Fig. 4.9. Both filters
converged in approximately 200 cycles, and managed to detect the known ∆θTDC ,
see Fig. 4.10. Moreover, the filters converged to slightly different final states in
Fig. 4.10, where the EKF converged to a higher C 2 value.

4.6 Discussion
For the choice of 250 particles, the PF was slower than the EKF (see Fig. 4.5).
The EKF is the minimum-variance unbiased state estimator if the system is linear
and perturbed with Gaussian noise, which could explain the higher convergence
rate in Fig. 4.5. The PF was however less sensitive to model-parameter errors and
had comparable RMSE in stationarity. Here, N p = 250 was chosen so that the
filters would have comparable computation times, and it is possible that the PF
performance would have been improved with an increased N p . The effects of N p
and the re-sampling criterion on PF performance were not addressed here and is
suggested future work.
The suggested parameter-estimation framework could easily be extended
or modified to cover other model parameters. It should however be kept in
mind that observability or identifiability might be degraded or lost when many

91
Chapter 4. Heat-Release Analysis

Table 4.2 Sensitivity to model-parameter errors. The filters were more sensitive
to errors in p in and r c and not as sensitive to other parameter errors. The PF was
overall less sensitive to model errors compared to the EKF. The units for θ∆TDC ,
C 2 and T w are CAD, m/(sK) and K.

EKF PF

p in error, 1.3 ± 0.065 bar

θ∆TDC = 0 ± 0.48 (0.38) θ∆TDC = 0 ± 0.3 (0.33)
C 2 = 0.0032 ± 0.003 (0.0024) C 2 = 0.0032 ± 0.0014 (0.0012)
T w = 497 ± 198 (151) T w = 497 ± 100 (95)

Tin error, 303 ± 20 K

θ∆TDC = 0 ± 0.17 (0.14) θ∆TDC = 0 ± 0.12 (0.12)
C 2 = 0.0032 ± 0.0003 (0.0001) C 2 = 0.0032 ± 0.00015 (0.0001)
T w = 497 ± 65 (45) T w = 497 ± 40 (25)

λ error, 2 ± 0.2
θ∆TDC = 0 ± 0.07 (0.09) θ∆TDC = 0 ± 0.05 (0.07)
C 2 = 0.0032 ± 0.00045 (0.0004) C 2 = 0.0032 ± 0.0004 (0.0003)
T w = 497 ± 12 (10) T w = 497 ± 7.5 (7.5)

r c error, 1.6 ± 0.8

θ∆TDC = 0 ± 0.65 (0.55) θ∆TDC = 0 ± 0.5 (0.51)

C 2 = 0.0032 ± 0.003 (0.003) C 2 = 0.0032 ± 0.0017 (0.0012)
T w = 497 ± 179 (138) T w = 497 ± 90 (70)

Tc error, 333 ± 20 K

θ∆TDC = 0 ± 0.01 (0.0) θ∆TDC = 0 ± 0.02 (0.025)

C 2 = 0.0032 ± 0.00005 (0) C 2 = 0.0032 ± 0 (0.0001)
T w = 497 ± 7 (4) T w = 497 ± 9 (10)

Q ctot error, 4 × 103 ± 200 J

θ∆TDC = 0 ± 0.15 (0.1) θ∆TDC = 0 ± 0.13 (0.12)
C 2 = 0.0032 ± 0.001 (0.0011) C 2 = 0.0032 ± 0.0012 (0.0013)
T w = 497 ± 30 (17) T w = 497 ± 0 (7)

92
 4.6 Discussion

Experimental Convergence
6,000
Q initial
EKFfinal
5,000
PFfinal
Q ctot
4,000

3,000
Q c [J]

2,000

1,000

−1,000
−100 −80 −60 −40 −20 0 20 40 60 80 100 120
θ [CAD]

Figure 4.9 Filter Q c -output convergence for the parameters presented in

Fig. 4.10. The EKF (black) and the PF (red) converged to slightly different Q c when
1,2
initiated at x̂ 0 (black, dashed). The black dash-dotted line indicates Q ctot com-
puted from fuel-flow measurements.

heat-release parameters are to be estimated simultaneously [Eriksson, 1998].

Moreover, it could be beneficial to include additional measurements from wall
temperature, rail pressure and λ sensors, along with sensor-uncertainty charac-
teristics. Model assumptions for Q c could also be developed further. For instance,
by making monotonic Q c more probable or by obtaining a more accurate Q c
noise model from experimental data.

Table 4.3 Operating point for experimental evaluation.

Nspeed [rpm] 1200 Tin [◦ C] 40

p IMEPn [bar] 10 λ [-] 1.8
p rail [bar] 800 Q ctot [J] 4680

93
Chapter 4. Heat-Release Analysis

Experimental Convergence
4
PF
EKF
θ TDC [CAD]
2

−2
0 50 100 150 200 250 300 350 400
cycle [-]

·10−2
1

0.8
C 2 [m/(sK)]

0.6

0.4

0.2

0
0 50 100 150 200 250 300 350 400
cycle [-]

800

600
T w [K]

400

200
0 50 100 150 200 250 300 350 400
cycle [-]

Figure 4.10 Filter convergence with respect to 400 cycles of experimental data
from the operating point in Table 4.3. The EKF (black) and the PF (red) have com-
parable performance but converged to slightly different state estimates. Both fil-
ters managed to detect a significant top-dead-center offset close to 1 CAD.

94
 4.7 Conclusions

4.7 Conclusions
A statistical framework for estimation of unknown heat-release model para-
meters was introduced in this chapter. Within this framework, the EKF and
the PF both seem to be feasible options for on-line estimation. The simulation
results showed that both filters were consistent in converging to the correct pa-
rameter values. The relation between assumed model and heat-release noise
variance determined a trade-off between convergence rate and steady-state
RMSE, and could be used as a tuning parameter. An assumed stochastic
accumulated-heat-release offset showed to be crucial when the injected fuel
energy varied. In reality, such variations are present due to common-rail pres-
sure fluctuations. The model-error-sensitivity results in Table 4.2 indicated that
the filters were more sensitive to intake-pressure and compression-ratio errors,
compared to other parameter errors. The model-error sensitivity was also found
to be dependent on the assumed heat-release noise variance. Furthermore, both
filters showed consistent convergence from different initial states with respect to
experimental data and manually tuned filter parameters, see Figs. 4.9 and 4.10.

95
5
Experimental Setup

5.1 The Scania D13 Engine

All experiments were performed on a Scania D13 six-cylinder heavy-duty die-
sel engine in the combustion-engine lab at Lund University, see Fig. 5.1. Eng-
ine specifications are given in Table 5.1. The original gas-exchange system
was extended with an additional water-cooled low-pressure EGR path and a
water-cooled air path prior to the intake manifold. The engine was boosted with
a fixed-geometry turbocharger.

Fuel
The fuel used was a mixture of 80 volume % gasoline and 20 volume % n-heptane.
This ratio was chosen based on previous results showing that a fuel octane num-
ber around 80 could be used over a wide range of engine operating points [Ma-
nente, 2010b]. The fuel was mixed together with an Infimeum fuel lubricant to
increase the lifetime of the fuel-injection system which was developed for con-
ventional diesel fuel.

Table 5.1 Engine Specifications

Total displaced volume 12.74 dm3

Number of cylinders 6
Stroke 160 mm
Bore 130 mm
Connecting rod length 255 mm
Compression ratio 18:1
Valves per cylinder 4
Maximum Power 360 kW

96
 5.2 Instrumentation

Figure 5.1 Scania D13 engine, located in the combustion-engine lab at Lund
University. Figure courtesy of Nhut Lahm.

5.2 Instrumentation
Actuation
Fuel Injection The fuel-injection system was a production extra-high
pressure-injection (XPI) common-rail system with solenoid injectors. The
common-rail pressure was regulated with an inlet-metering valve, positioned
prior to a fuel pump used to elevate the pressure in the common rail volume.
Fuel-injection timings and durations were determined by current pulses sent to
the injectors. Current-pulse timings and durations were set from the LabVIEW
control system and actuated with Drivven direct-injection drivers. A more detai-
led description of the fuel-injection system is given in [Källkvist, 2011].
Gas Flow The engine was equipped with two cooled EGR loops, located before
and after the turbine. EGR flows were regulated with two valves. Two valves posi-
tioned prior to the intake manifold were used to regulate the intake temperature
by adjusting the flow over an intercooler. A back-pressure valve positioned after
the tubine was used to create back pressure for sufficient EGR flow. Servo motors
used for valve actuation were controlled from the LabVIEW control system and
actuated with Drivven drivers. Valve locations are marked in Fig. 5.2.
Engine Speed The engine speed was controlled with an ABB M2BA electrical
motor with a rated power of 355 kW. The motor reference speed was adjusted
manually from the engine control room.

97
Chapter 5. Experimental Setup

air

intercooler
θcool ṁ a
T
p, T
p

θhot

EGR p, T
EGR

p p, T p, T
p, T
T p, T
p
p, T
T compressor θLP
p
p, T θHP
T
p
p turbine
p, T
T
exhaust manifold
intake manifold

λ
p
p, T
T p, T

p
p, T
T
ṁ f
CO2 p

θBP
M,N
fuel em

exhaust
- valve
electrical motor

Figure 5.2 A schematic illustration of the engine configuration with locations of

sensors and valves. Sensor symbols: p - pressure, T - temperature ṁ f - fuel-mass
flow, ṁ a - air-mass flow, λ - air/fuel ratio, M - torque, N - engine speed, em -
emissions. Valves are denoted θx .

Sensing
Sensor locations are marked in Fig. 5.2.

Sampling Crank-angle based sampling was enabled through a Leine & Linde
encoder emitting 5 pulses every CAD which triggered sampling of cylinder pres-
sure, engine torque and injector current. Other sensor signals were sampled
every engine cycle.

In-cylinder pressure sensors The in-cylinder pressure was measured with

water-cooled Kistler 7061B piezo-electrical pressure tranducers. The cylinder-pressure
signal was sampled every 0.2 CAD with the crank-angle encoder.

98
 5.2 Instrumentation

Common-rail pressure sensor The common-rail pressure was measured with

pressure sensor mounted to the common-rail volume.

Pressure and temperature sensors Pressures and temperatures were measured

at various locations in the gas-exchange system. Keller PAA-23S absolute press-
ure sensors with response times of milliseconds were mounted accordingly:

• intake manifold close to the intake valves of cylinders 1-6.

• exhaust manifold.

• after the turbine.

• before and after the compressor.

• after the thermal-management system.

• after the low-pressure EGR valve.

• after the high-pressure EGR valve.

Temperatures were measured with K-type thermocouples with response times of

seconds. These sensors were mounted accordingly:

• intake manifold close to the intake valves of cylinders 1-6.

• exhaust manifold at the exhaust valves of cylinders 1-6.

• after the turbine.

• before and after the compressor.

• after the thermal-management system.

• after the low-pressure EGR valve.

• after the high-pressure EGR valve.

Torque sensor A force sensor integrated in the electrical motor was used to me-
asure engine torque.

Engine Speed Engine speed was obtained from the internal speed measure-
ment of the electrical motor.

Fuel flow A Bronkhorst mini CORI-FLOW M15 mass-flow meter mounted prior
to the fuel system was used to measure the fuel-mass flow.

Air flow A Bronkhorst hot-film air-mass flow meter mounted prior to the com-
pressor was used to measure the air-mass flow.

99
Chapter 5. Experimental Setup

λ sensor A broadband λ sensor mounted after the turbine measured the ex-
haust oxygen concentration.
Emissions Intake and exhaust CO2 and exhaust NOx , HC, CO and O2 levels
were measured with an AVL AMA i60 exhaust-measurement system. Soot levels
were measured with an AVL micro soot sensor measurement unit.

5.3 Control-System Architecture

Hardware
The engine was controlled with a real-time system, consisting of a NI PXIe-8135
embedded controller with a 2.3 GHz quad-core processor, and NI PXI-7854/7854
R which is a multifunction reconfigurable I/O with Virtex 5-LX110/LX30 FPGAs.
The FPGAs were used as configurable hardware for flexible AO / DIO and AD
acquisition, triggered by the crank-angle encoder. The ADC sampled analog sig-
nals with a 16-bit resolution. A user interface was run on a separate host PC with
a Windows 7 operating system which communicated with the real-time system
over TCP/IP.

Software
The engine control system was programmed in LabVIEW which is a graphical
programming environment developed by National Instruments. The software
was originally developed by Borgquist for his thesis work [Borgquist, 2013].
Real-time heat-release analysis and controller computations were executed
by the real-time PXI system. Computations were done using floating point arith-
metic, and most of them were done in LabVIEW MathScript RT Module nodes
inside timed loops, triggered every engine cycle. PI controllers were implemen-
ted using the LabVIEW PID advanced VI and QPs were solved using the LabVIEW
quadratic programming VI. The quadratic programming VI had functionalities
useful for model predictive control implementation such as initialization, warm
start of active constraints, various stopping criteria and error flags when feasible
solutions were not found. The user interface was also programmed in LabVIEW.

Signal Processing
The in-cylinder pressure was measured with piezo-electric transducers. This
measurement technique has high cut-off frequency, good linearity and handling
of the environment inside the combustion chamber. The signal given by these
sensors is on the form
p meas = kp + ∆p (5.1)
where p meas is the sensor signal, p the actual pressure, k the sensor conversion
factor and ∆p the sensor offset. Methods for determining k and ∆p were pre-
sented by Randolph [1990]. In this work, k were known from sensor calibration

100
 5.3 Control-System Architecture

and ∆p was determined by referencing the cylinder pressure at intake valve clo-
sing (IVC) to the measured intake-manifold pressure p in . High-frequency con-
tent in p meas during combustion and expansion was attenuated using a digital
zero-phase filter.

101
6
Proportional-Integral
Combustion-Timing Control

This chapter investigates how a proportional-integral (PI) controller should be

designed for robust and noise insensitive combustion-timing control through
injection-timing actuation. The aim of the chapter is to provide physical under-
standing of the combustion-timing control problem and to illustrate how para-
meters such as fuel reactivity and engine load affect controller performance.
Combustion-timing feedback is motivated by the sensitivity of partially pre-
mixed combustion, where the increased importance of autoignition reactions
leads to an increased combustion-timing variability. Disturbances in injected
fuel amount and EGR flow could either result in a too early combustion ti-
ming with excessive in-cylinder temperatures and rapid combustion rates, or a
too late combustion timing with incomplete combustion and resulting hydro-
carbon emissions. This sensitivity has previously been described in [Ekholm et
al., 2008; Yin et al., 2015; Henningsson, 2012; Li et al., 2016]. Additional motives
for combustion-timing feedback are hardware variation and aging and the in-
creased variation in fuel properties due to the introduction of different types of
biofuels.
In-cylinder pressure measurements allow for detection of the combustion ti-
ming, indicated by θ50 , as described in Chapter 4. Detected θ50 can be regulated
cycle-by-cycle by varying the injection timing θSOI as illustrated in Fig. 6.1, where
a controller determines θSOI from the θ50 set-point deviation e. Closed-loop θ50
control of this type was previously used in [Shaver et al., 2004; Bengtsson et
al., 2004; Willems et al., 2010].
The proportional-integral (PI) controller is the most common solution to
practical closed-loop control problems and is attractive because of its simpli-
city. Furthermore, the problem of designing a PI controller illustrates the inhe-
rent limitations and complexities of combustion-timing feedback, which gives
valuable insights for other controller-design approaches.

102
 Chapter 6. Proportional-Integral Combustion-Timing Control

Disturbance

r
θ50 P e θSOI θ50
Controller Engine

−1

Figure 6.1 The combustion-timing feedback loop. The indicator for combustion
timing θ50 is obtained from the measured in-cylinder pressure and heat-release
r −θ caused by,
analysis. The controller varies θSOI to counteract the error e = θ50 50
e.g., changes in fuel-amount or intake conditions.

The PI controller considered is given by

θSOI (k + 1) = k p e(k) + I (k)

(6.1)
I (k) = I (k − 1) + k I e(k − 1)

where the injection timing of the following cycle θSOI (k + 1) is determined by the
previous-cycle error
r
e(k) = θ50 (k) − θ50 (k) (6.2)
multiplied with the proportional gain k p , and the integral term I (k), which is the
sum of previous errors, scaled with the integral gain k I . The integral term is intro-
duced to bring e to zero in steady state.
When introducing feedback, the controller has to be robustly designed to en-
sure closed-loop stability. It is also important that the controller does not en-
hance stochastic cycle-to-cycle variation. In PI-controller design, this is done by
carefully deciding the controller gains k p and k I . The problem of deciding k p and
k I is investigated in this chapter by evaluating controller performance through
simulation. Simulation allows for evaluation of a large number of gain combina-
tions at different engine loads.
The effect of fuel reactivity on controller design is addressed by evaluating
controller performance for different primary reference fuels (PRF). A PRF is
a mixture of n-heptane and iso-octane, where the PRF number indicates the
iso-octane volume percentage. Primary reference fuels are commonly used as
reference in engine research and are used to determine the octane number of
a fuel. The octane number (ON) is a measure of the fuel resistance to autoigni-
tion which increases with ON, and is defined by the PRF value needed to provide
equivalent autoignition properties.
The presented controller-gain evaluation provides gains that simultaneou-
sly maximize attenuation of θ50 disturbances and fulfill constraints on robust-

103
Chapter 6. Proportional-Integral Combustion-Timing Control

ness and noise sensitivity. The evaluation also investigates the trade-off between
these performance requirements. The optimization-based approach to PI con-
troller design was inspired by works presented in [Hast et al., 2013; Garpinger
and Hägglund, 2015], where understanding and rules of thumb for PI controller
design were found through optimization.
The chapter is outlined as follows: The model used for controller evaluation
is introduced in Sec. 6.1. The steady-state relation between θSOI and θ50 is stu-
died in Sec. 6.2. Section 6.3 presents the criteria used for controller evaluation.
Robust and noise-insensitive controllers, found through simulation of transient
and steady-state operation for different PRFs are presented in Sec. 6.4, together
with an analysis of the results. Finally, conclusions are given in Sec. 6.5.

6.1 Modeling
The in-cylinder state and wall-surface temperature were modeled using the
zero-dimensional model presented in Sec. 2.2. Constant-volume combustion
and static gas-exchange boundary conditions were assumed to speed up com-
putations. A detailed description of this model is given in [Widd et al., 2008]
where it was shown to successfully predict experimental θ50 and p IMEPg . The
ignition-delay τ was computed using the model M4, presented in Sec. 2.4

ζ(T,PRF) Λ(T,PRF)
τ = φα(T,PRF) p β(T,PRF) xO e (6.3)
2

This is a PRF correlation, calibrated from constant-volumes simulation data over

a wide range of engine-relevant operating conditions. The model was presen-
ted in [Delvescovo et al., 2016] and was able to predict θ50 in experimental HCCI
operation. The start of combustion θSOC was computed using the Livengood-Wu
integration criterion [Livengood and Wu, 1955]
Z θSOC 1
dt = 1 (6.4)
θSOI τ

The model of the closed engine cycle was used in closed-loop simulation ex-
periments, for which the model outputs θ50 and p IMEPg were regulated on a
cycle-to-cycle basis using θSOI and the injected fuel energyf Q ctot . Model para-
meters used are presented in Table 6.1.

6.2 Steady-State Characteristics

Fuel injection provides direct control of the combustion processes. This means
that the combustion timing can be adjusted from one engine cycle to the
next. There is also weak cycle-to-cycle dependence due to residual gas and

104
 6.2 Steady-State Characteristics

Table 6.1 Model parameters used in simulation. Three different convection co-
efficients were used during compression (compr.), expansion (exp.) and gas ex-
change (gas ex.).

Cylinder Geometry Heat-Transfer

Comp. ratio [-] 18 h c (comp.) [W/(m2 K)] 250

Vd [m3 ] 2.1 × 103 h c (exp.) [W/(m2 K)] 500
Bore [mm] 130 h c (gas ex.) [W/(m2 K)] 250
Stroke [mm] 160 Tc [K] 333
IVC [CAD] -151 m c c p [J/K] 1150
EVO [CAD] 146 k c [W/(m2 K)] 45
L c [m] 0.025

slow wall-temperature dynamics. These effects will be discussed in the fol-

lowing sections. The strong direct effect between θSOI and θ50 makes the system
steady-state characteristics important in the controller design.
The model steady-state input-output relation between θSOI and θ50 is pre-
sented in Fig. 6.2 for PRF20 (red) and PRF100 (blue). The relation is represented
as bands, generated from a range of engine loads. Load was varied by increa-
sing the fuel energy Q ctot from 2000 to 6000 J, at a relative air-fuel ratio λ = 2, and
intake temperature Tin = 70 ◦ C. These values were chosen to resemble typical
mid-to-high load conditions in the author’s engine lab, with the exception of the
high Tin .
The system exhibits a nonlinear input/output behavior. For a given Q ctot , there
is a limited interval of obtainable θ50 , and an interval of θSOI for which θ50 is
controllable. It is within these intervals the PI controller should operate. For late
θSOI , θ50 is excessively delayed which indicates that ignition never occurs, i.e., the
charge misfires. This happens when insufficient time and reactivity during the
expansion stroke result in an unfulfilled ignition condition, see (6.4). For early
θSOI , the system gain approaches zero due to low reactivity during the early com-
pression stroke, which results in poor controllability.
There is a clear difference between the two fuels. The obtainable θ50 interval
is narrower with a higher PRF value due to the decreased fuel reactivity. With
Q ctot = 2000 J, it is not possible to obtain θ50 < 6 CAD for PRF80 due to insuffi-
cient reactivity during compression. There is also a visible difference in Q ctot sen-
sitivity for the two fuels. In order to obtain similar θ50 and conduct a comparable
controller-evaluation for the different fuels, Tin was adjusted as a function of PRF.
This adjustment was done to ensure that θ50 = 5, an efficient θ50 set point, could
be obtained from PRF0 to PRF100 at the lowest Q ctot . Found Tin values are presen-
ted in Table 6.2. A variable compression ratio or variable valve timings are other
possible adjustments for fuel-reactivity compensation. These solutions would,

105
Chapter 6. Proportional-Integral Combustion-Timing Control

PRF20 & PRF80, λ = 2, Q ctot = 2000 to 6000 J, Tin = 70◦ C

30
PRF = 20
PRF = 80
Misfire →

20
Q ctot = 2000 J
θ 50 [CAD]

10 ← Not Controllable

Q ctot = 6000 J
0

−10
−40 −30 −20 −10 0 10 20
θ SOI [CAD]

Figure 6.2 Steady-state θ50 / θSOI relation for PRF20 and PRF80 with fuel ener-
gies from 2000 to 6000 J. The Q ctot sweep generates a band of θ50 / θSOI relations.
For late θSOI , θ50 is excessively delayed which means that ignition never occurs.
Controllability is reduced when θSOI decreases.

Table 6.2 Relation between PRF and Tin in order to obtain θ50 = 5 CAD with
λ = 2 and Q ctot = 2000 J.

PRF [-] 0-10 20 30 40 50 60 70 80 90 100

◦
Tin [ C] 50 53 58 62 67 70 73 76 78 80

however, be more demanding on engine hardware, as Tin could be varied using

fast thermal management [Haraldsson, 2005]. The Tin adjustment gave the θ50 /
θSOI characteristics in Fig. 6.3, where the layered bands represent θ50 / θSOI rela-
tions from PRF0 (dark red) to PRF100 (dark blue). Now, θ50 ∈ [5, 15] are obtainable
for all fuels and Q ctot . There are however still differences in Q ctot sensitivity, which
decreases from PRF0 to PRF100 in the θSOI interval for which θ50 is controllable.
There is also a shape difference between for the different fuels. The gain chan-
ges more steeply as θSOI is delayed prior to the misfire region for higher PRF val-
ues, whilst lower PRF values show a more linear trend. Figure 6.4 shows how this

106
 6.2 Steady-State Characteristics

PRF0-100, λ = 2, Q ctot = 2000 to 6000 J, adjusted Tin

30
PRF0
25 PRF100

20
Q ctot = 2000 J
15
θ 50 [CAD]

10

Q ctot = 6000 J
0

-5

-10
−40 −30 −20 −10 0 10 20
θ SOI [CAD]

Figure 6.3 Steady-state θ50 / θSOI relation with adjusted Tin are presented as
layered bands. The lower red band corresponds to PRF0 and the upper blue band
to PRF100.

difference can be explained by how τ varies as a function of θ. Higher PRF values

have a narrower range with shorter τ, and exhibit a different τ behavior close to
TDC.
This difference is caused by the negative-temperature coefficient (NTC) be-
havior of low-value PRFs. NTC means that the reaction rate decreases with incre-
asing temperature, and is the result of an increase in relative importance of ter-
minating reaction paths over branching reaction paths, which slows down overall
reaction rates [Curran et al., 2002]. The ignition delay therefore increases with in-
creasing temperature for certain operating conditions. For the lower value PRFs,
τ starts to increase with CAD during the compression stroke, and then decreases
again as θSOI is delayed towards top-dead center. The NTC behavior is then re-
peated during the expansion stroke and gives an overall more constant τ close to
TDC.
In order to ensure adequate PI-controller performance, it is necessary to limit
θSOI in an interval where θ50 is controllable whilst avoiding misfire. Misfire leads
to irregular power output and greatly increases fuel consumption. Too early in-
r
jection timings, as a result of infeasible θ50 , leads to controller wind up. The re-
gion where 0.1 < ∂θ50 /∂θSOI < ∞ is represented by the shaded areas in Fig. 6.5 for
PRF0 and PRF100. The region is narrower for low Q ctot and high PRF values. This

107
Chapter 6. Proportional-Integral Combustion-Timing Control

Ignition delay, τ, used in the autoignition integral

10

6
τ [ms]

2 PRF0
PRF100
NTC behavior

0
−30 −20 −10 0 10 20 30
θ [CAD]

Figure 6.4 Ignition delay τ used in the autoignition criterion in (2.69) as a func-
tion of θ with adjusted Tin according to Table 6.2. The low-value PRFs exhibit a
NTC behavior during compression and expansion. Note that both temperature
and PRF value are varied for the different curves, and that the NTC behavior is
dependent on both parameters. It was decided to study NTC behavior after Tin
adjustment since this would be necessary prior to engine operation to get satis-
factory operation for the high PRF values.

implies that a PI controller alone is not sufficient. A controller also has to be able
r
to detect misfire, and if θ50 is infeasible to avoid wind up. The controller should
r
then take action by adjusting θ50 or limit θSOI . Design of such a controller is not
covered here.

6.3 Controller Evaluation

This section presents the controller-evaluation method used. The approach ad-
opted was to evaluate PI-controller performance over a grid of controller gains.
Optimal controller gains were computed by evaluating accumulated tracking er-
ror, robustness to instability and θSOI variation due to stochastic disturbances.
Tracking error was computed during a simulated test cycle where changes in
r r
θ50 , Tin and p IMEPg were made. Cycle-to-cycle variation in θSOI was evaluated
in steady-state where the model was exposed to stochastic disturbances.

108
 6.3 Controller Evaluation

Controllable Region, 0.1 < ∂ θ 50 / ∂ θ SOI < ∞

PRF0
PRF100
5,000
Q ctot [J]

4,000

3,000

−35 −30 −25 −20 −15 −10 −5 0 5

θ SOI [CAD]

Figure 6.5 The region where 0.1 < ∂θ50 /∂θSOI < ∞. In order for the controller
to perform satisfactorily, θSOI should be limited within this region. Note that the
region is narrower for low load and PRF100.

Load Control
The gross indicated mean effective pressure p IMEPg was also controlled during
simulation. This was done using a PI controller that adjusted Q ctot for tracking of
r
a set point p IMEPg . The PI-controller gains were tuned for a response time of 10
cycles. Robustness and measurement-noise sensitivity were therefore evaluated
with respect to the resulting multiple input/output system to account for cross
coupling from Q ctot to θ50 , and from θSOI to p IMEPg .

Disturbance Rejection
r
The θ50 -controller objective is to follow θ50 changes whilst attenuating the effects
of disturbances, see Fig. 6.1. The controller ability to fulfill this objective was me-
asured by computing the discrete-time integrated absolute error (IAE)

N
X
IAE = |e(k)| (6.5)
k=1

during the simulation experiments, where N is the number of cycles. The IAE is a
commonly used measure for controller evaluation [Åström and Hägglund, 2006].

109
Chapter 6. Proportional-Integral Combustion-Timing Control

Here, it will mainly penalize transient control error since the controllers exhibit
error-free tracking in steady state due to integral action.

Robustness
It is important that the controller is stable around the set point. This can be en-
sured by having sufficient stability margins. For linear discrete-time systems, ro-
bustness is captured by the sensitivity function S and the complementary sensi-
tivity function T [Zhou and Doyle, 1998]

S(z) = (I + P (z)C (z))−1

(6.6)
T (z) = P (z)C (z)(I + P (z)C (z))−1

Here, z ∈ C and I denotes the identity matrix, P is the pulse-transfer function of

the system to be controlled and C is the controller in feedback interconnection.
Bounds on the H∞ -norms of S and T can be introduced as guarantees for robus-
tness
M s = ||S(e i ω )||∞ ≤ κM s
M t = ||T (e i ω )||∞ ≤ κM t (6.7)
ω ∈ [0, π]
An elaborate description of robust controller design can be found in [Zhou and
Doyle, 1998]. Norm values ranging from 1.2 to 2 correspond to gain margins from
0 to 2 and phase margins between 49◦ to 29◦ [Åström and Hägglund, 2006]. Here,
κM s = κM t = 1.4 were introduced as upper bounds.
The norms M s and M t were computed through model linearization. Line-
arization was done with step-response analysis at multiple stationary points
along the test-cycle trajectory, indicated by the symbols in Fig. 6.7. The obtai-
ned worst-case M s and M t were then used in (6.7). The transfer function of the
linearized model were on the form
à !
P 11 (z) P 12 (z)
P (z) =
P 21 (z) P 22 (z)
(6.8)
k iwall
j
(1 + a wall )z −1
P i j (z) = k idj + k ires
j z
−1
+
1 + a wall z −1

Here, P 11 is the transfer function from θSOI to θ50 , P 22 is the transfer function
from Q ctot to p IMEPg and P 12 , P 21 represent the corresponding cross-coupling
dynamics. Furthermore, k idj is a direct gain, k ires
j
models the one-cycle delayed
residual-gas effect, and a wall and k iwall
j
determine the time constant and gain of
the wall-temperature dynamics. Step responses for P are presented in Fig. 6.6, for
PRF0 at p IMEPg = 6 and θ50 = 12.

110
 6.3 Controller Evaluation

P 11 P 12
−1
residual gas
1.4 −1.2
θ 50 [CAD]

θ 50 [CAD]
−1.4
1.2 −1.6
wall temperature
−1.8
direct effect
1 −2
0 100 200 300 400 0 100 200 300 400
cycle [-] cycle [-]

P 21 P 22
0 3
p IMEPg [bar]

p IMEPg [bar]
−0.05
2
−0.1
1
−0.15

−0.2 0
0 100 200 300 400 0 200 400 600 800
cycle [-] cycle [-]

Figure 6.6 Step response for (6.8). Input steps of θSOI = 1 CAD and Q ctot = 1000
J are applied at cycle 1. The P 11 subdiagram indicates the direct effect, the
residual-gas effect and the wall-temperature dynamics, respectively.

The 2 × 2 controller is given by

à !
C 1 (z) 0
C= (6.9)
0 C 2 (z)

where C 1 and C 2 are θ50 and p IMEPg PI controllers on the form

k Ii
C i (z) = z −1 (k pi + ) (6.10)
z −1

Where z −1 is the delay from measurement to actuation. The reason for desig-
ning a decoupled controller, and for designing C 2 given C 1 , was because of weak
coupling (see P 12 and P 21 in Fig. 6.6) and the convenience of designing a load
controller with respect to other aspects than θ50 .

Noise Sensitivity
It is necessary to avoid excessive θSOI variation due to stochastic cycle-to-cycle
variation. This requirement was formulated as a constraint on the steady-state

111
Chapter 6. Proportional-Integral Combustion-Timing Control

θSOI standard deviation

σθSOI ≤ κσ (6.11)

Stochastic cycle-to-cycle variation was evaluated by introducing Gaussian-noise

disturbances on Q ctot , Tin and p in . It was decided to set the upper bound κσ = 0.25
to maintain θSOI within ± 1 CAD in steady state.

Optimization Problem
In summary, the design problem is described by the following optimization pro-
blem from PRF0 to PRF100

N
X
minimize |e(k)| (6.12)
kp , kI k=1

subject to M s ≤ 1.4
M t ≤ 1.4
σθSOI ≤ 0.25

The optimization problem was solved by simulating the system model and eval-
uating (6.12) over a grid of controller gains. This was not only done to find opti-
mal gains but also to investigate cost function and constraint characteristics as a
function of k I , k p , Q ctot and PRF.

6.4 Results
The optimization problem (6.12) was solved by evaluating a grid of gains k p , k I ∈
{0.05, 0.1, . . . , 1} and PRFs ∈ {0, 10, . . . , 100} during two simulation experiments:

1. A transient cycle consisting of 750 engine cycles for which disturbances in

r r
Tin (±5◦ C ), and changes in θ50 (±6 CAD) and p IMEPg (±10 bar) were made
to evaluate M s , M t and IAE.

2. A steady-state noise-sensitivity experiment consisting of 8000 engine cy-

r r
cles at p IMEPg = 5 and 15 bar with θ50 = 6 and 12 CAD. Gaussian-noise
disturbances were applied to Tin , p in and Q ctot , with standard deviations
σTin = 1 ◦ C, σQ ctot = 50 J and σp in = 0.025 bar.

The relative air-fuel ratio λ was set equal to 2 by adjusting p in , whereas Tin was
adjusted as a function of PRF according to Table 6.2.

112
 6.4 Results

Optimal θ 50 trajectories for PRF0 to PRF100

20
r
Tin disturbance p IMEPg change
r
θ50 change
10 θ50
θ 50 , θ SOI [CAD]

0
θSOI

−10
PRF0
PRF100
−20
100 200 300 400 500 600 700
cycle [-]

Figure 6.7 Optimal θ50 and θSOI trajectories from PRF0 to PRF100. Changes in
r
Tin were applied at cycles 50, 100, 200, 250, 400, 450, 550 and 600, p IMEPg chan-
ges were applied at cycles 350 (increase) and 700 (decrease), and θ50r was varied at
r
cycles 150, 300, 350, 500, 650 and 750. During the p IMEPg r was also chan-
steps, θ50
ged in order to increase θSOI variation. The symbols indicate the points of linea-
rization where M s and M t where computed, see Table 6.3 for static-gain values.
When comparing PRF performance, the difference in θ50 was not as significant as
the difference in θSOI for the different PRFs. It can be seen that changes in p IMEPg
r resulted in greater θ
and θ50 SOI variation for lower PRF values, and that the θSOI
response was more comparable during Tin changes.

Optimal Gains
Optimal transient θ50 and θSOI trajectories with respect to (6.12) are presented in
Fig. 6.7, where optimal time constants of the closed loops are within 10 cycles for
all disturbances and set-point changes. When comparing different PRFs, it can
be seen that the difference in θ50 was not as significant as the difference in θSOI .
r
Load and θ50 variation gave greater θSOI variation for low-value PRFs, whilst θSOI
variation was comparable for different PRFs during Tin changes.
Linearization points for which robustness was evaluated are indicated by the
symbols 5, 2, ° and ∗ in Fig. 6.7. Computed static gains with respect to these
points are presented in Table 6.3 for PRF0 and PRF100. It can be seen that the P 11
gains were higher for late θ50 and PRF100, and that the interaction from Q ctot to
r
θ50 (P 12 ) was higher at low p IMEPg .
Optimal steady-state performance is presented in Fig. 6.8. The θSOI standard
deviation σθSOI was higher for the low-load operating points (cycles 1-500), σθSOI
was also higher for lower PRF values.

113
Chapter 6. Proportional-Integral Combustion-Timing Control

Table 6.3 Static gains at the linearization points. Units are given by [-], [CAD/kJ],
[bar/CAD] and [bar/J] for P 11 , P 12 , P 21 and P 22 . Note the following trends: higher
P 11 gains at late θ50 and PRF100 and a larger interaction from Q ctot to θ50 at low
p IMEPg .

PRF0

P5 P2
r r r r
p IMEPg = 5 bar, θ50 = 6 CAD p IMEPg = 5 bar, θ50 = 12 CAD
à ! à !
0.35 −0.005 1.35 −0.007
0.0033 0.0029 −0.07 0.003

P∗ P°
r r r r
p IMEPg = 15 bar, θ50 = 6 CAD p IMEPg = 15 bar, θ50 = 12 CAD
à ! à !
0.9 −0.001 1.35 −0.0015
−0.056 0.0029 −0.16 0.0029
PRF100

P5 P2
r r r r
p IMEPg = 5 bar, θ50 = 6 CAD p IMEPg = 5 bar, θ50 = 12 CAD
à ! à !
0.7 −0.0025 1.85 −0.004
−0.0011 0.0028 −0.1 0.0029

P∗ P°
r r r r
p IMEPg = 15 bar, θ50 = 6 CAD p IMEPg = 15 bar, θ50 = 12 CAD
à ! à !
1 −0.0005 1.4 −0.001
−0.065 0.0029 −0.16 0.0028

Finally, optimal controller gains are presented in Fig. 6.9 as a function of PRF.
Both k p and k I decreased with PRF and k p was slightly lower than k I . The M s
value at the 2-linearizing point in Fig. 6.7 was always found to be the constraint
limiting controller-gain magnitudes. At this point, the system gain ∂θ50 /∂θSOI
was highest among the linearization points, see Table 6.3. This partial derivative
also increased with PRF, which explains the trend in Fig. 6.9. These trends could
have been anticipated by studying Fig. 6.3, where ∂θ50 /∂θSOI increased with θ50 ,
PRF value, and decreased with Q ctot . The remainder of this chapter investigates
how the robustness and noise-sensitivity constraints vary with PRF.

114
 6.4 Results

Optimal Steady-State Performance for PRF0 to PRF100

20
PRF0
PRF100

10
σθSOI = 0.052
θ 50 , θ SOI [CAD]

σθSOI = 0.123
σθSOI = 0.044
0
σθSOI = 0.136
σθSOI = 0.111

−10 σθSOI = 0.223 σθSOI = 0.117

σθSOI = 0.180
−20
100 200 300 400 500 600 700 800 900 1,000
cycle [-]

Figure 6.8 Optimal steady-state performance from PRF0 to PRF100. The θSOI
standard deviation σθSOI was higher for the low p IMEPg points (cycles 1-500), and
for the lower PRFs. The presented data is a part of the steady-state experiment
consisting of 8000 cycles.

0.5
kp
kI
Controller Gains [-]

0.4

0.3

0.2
0 10 20 30 40 50 60 70 80 90 100
PRF [-]

Figure 6.9 Optimal controller gains as a function of PRF. The M s < 1.4 constraint
at the 2 operating point was consistently limiting the controller gains. The sys-
tem gain ∂θ50 /∂θSOI was highest at this point. Furthermore, ∂θ50 /∂θSOI increased
with PRF which explains the trend in this figure.

115
Chapter 6. Proportional-Integral Combustion-Timing Control

IAE and M t / M s Level Curves

1
IAE M s , M t = 1.4, ON = 0 M s , M t = 1.4, ON = 100
0
40 1.4
0 0
3
5 0
0.8 2
0

1.4
20
100
1. 4

1. 4
1.4
150

0.6
k I [-]

1. 4
150

1. 4
0.4 1.4
1.4
1.4

200

250
1.4
300
1.4

0.2
400

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

k p [-]

Figure 6.10 IAE (black), M s = 1.4 (solid) and M t = 1.4 (dashed) level curves as
a function of controller gains k p and k I , for PRF0 and PRF100. The IAE score de-
creased as k p and k I increased. The allowed gains for which M s , M t < 1.4 are
encircled by the colored lines, and optimal gains are indicated by circles.

Robustness
To illustrate robustness-constraint characteristics, IAE level curves (black, solid)
are presented in Fig. 6.10, together with level curves of M s = 1.4 (solid) and M t =
1.4 (dashed) as a function of controller gains for PRF0 and PRF100. The IAE score
decreased as k p and k I were simultaneously increased. The stability limits are
indicated by the steep increase in IAE in the upper-left and lower-right regions
of the figure. The constraint on M t was less restrictive than the constraint on M s ,
and the M s and M t constraints were overall less restrictive for PRF0. The limiting
r r
M s value was computed at the late θ50 and low p IMEPg operating point 2. The
PRF0 fuel also had visible M s and M t constraints for low controller gains due to
interaction with the p IMEPg loop. This was a result of the higher P 12 gain for PRF0.
These results can be explained by linear-systems analysis. With the

116
 6.4 Results

Nyquist Diagram
0.5

0
Imaginary Axis [-]

−0.5

−1

−1.5 P 11C 1 w/o dynamics

P 11C 1
Eq. (6.13)
−2
−0.5 −0.45 −0.4 −0.35 −0.3 −0.25 −0.2 −0.15 −0.1 −0.05 0
Real Axis [-]

Figure 6.11 Nyquist curves with optimal controller gains for PRF0 (red) and
PRF100 (blue), for (6.13) (solid), P 1C 1 (dashed), and P 1C 1 without residual and
wall-temperature dynamics (dotted). It can be seen that the simple model has the
smallest stability margin.

p IMEPg -loop closed, the θ50 open-loop transfer function is given by

¡ P 12C 2 P 21 ¢
P 11 − C1 (6.13)
1 + P 22C 2

Nyquist curves with k I = 0.35, k p = 0.3 for PRF0 (red) and PRF100 (blue) are pre-
sented in Fig. 6.11. This figure presents Nyquist curves for (6.13) (solid), P 1C 1
(dashed) and P 1C 1 without residual and wall-temperature dynamics (dotted). It
can be seen that the simplest open-loop transfer function has the smallest sta-
bility margin. The intuition behind this result is that both residual-gas dynamics
and the p IMEPg controller counteract the θSOI effect on θ50 , which decreases the
open-loop gain. Analysis of the simple system can therefore be used to compute
stability margins that are sufficient for the more detailed loops.
When omitting wall-temperature, residual-gas and p IMEPg -loop dynamics,
r
the linearized closed-loop pulse-transfer function, from θ50 to θ50 is given by

K C 1 (z)
Hcl (z) = (6.14)
1 + K C 1 (z)

117
Chapter 6. Proportional-Integral Combustion-Timing Control

Stability Region

1.4
← K = 0.9

1.2
←K =1

1
← K = 1.1

0.8
k I [-]

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
k p [-]

Figure 6.12 PI-gain stability region, computed using the critera in (6.16). The
stability region shrinks as K increases. The stability limits are similar to the IAE
level curves in Fig. (6.10).

where K = ∂θ50 /∂θSOI . Inserting the pulse transfer function for (6.10) into (6.14)
yields the characteristic polynomial for the closed-loop poles

z 2 + z(K k p − 1) + K (k I − k p ) (6.15)

By applying the Jury stability criterion [Jury, 1964], and assuming positive con-
troller gains, the conditions for stability is obtained

k p < k I /2 + 1/K
(6.16)
k p > k I − 1/K

The controller-gain stability region becomes narrower as K increases, see

Fig. 6.12.

Noise Sensitivity
Figure 6.13 presents IAE (black) and σSOI = 0.25 level curves for PRF0 and PRF100
as functions of k p and k I . Allowed gains for which σSOI ≤ 0.25 are within the red
and blue lines. The σSOI constraint was more restrictive for PRF0, and the most
restrictive σSOI constraint was found at late θ50 and low p IMEPg for both fuels, see
Fig. 6.8.

118
 6.4 Results

IAE and σθSOI Level Curves

1
IAE σθSOI = 25
0. 0.25, PRF = 0 σθSOI = 0.25, PRF = 100

0
10
00

5
0.2
4
0

0
30

15
0.8 0
25 10
0

2 00 100

0.25
5
0.2
0.25

0.25
0.6
k I [-]

0
0.2
5 15

0.25
150

0. 25
0.4
200

250
0.25

300
5
0.2 0.2
400

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

k p [-]

Figure 6.13 IAE (solid, black) and σSOI = 0.25 level curves for PRF0 and PRF100
as a function of k p and k I . The allowed gains for which σSOI ≤ 0.25 are within
the solid colored lines. The most restrictive σSOI = 0.25 constraints, for both fuels
were found at the late θ50 at the low-load operating points, see Fig. 6.8.

The θSOI standard deviation can be evaluated by studying the H2 -norm of the
transfer function S c

S c = C (z)(I + P (z)C (z))−1 (6.17)

which maps σθ50 to σSOI . The H2 -norm of S c amplifies σθ50 according to

σSOI = ||S c ||2 σθ50 (6.18)

With independent disturbances on Tin , Q ctot and p in , σθ50 can be approximated

through linearization, using the expression
sµ ¶2 µ ¶2 µ ¶2
∂θ50 ∂θ50 ∂θ50
σθ50 = σ2T + σ2 tot + σ2p in
∂Tin in ∂Q c Qc ∂p in

119
Chapter 6. Proportional-Integral Combustion-Timing Control

0 0

∂ θ 50 / ∂Q ctot [CAD/J]
∂ θ 50 / ∂Tin [CAD/K] −0.2
−5
−0.4

−0.6 −10

−0.8
5 10 15 5 10 15
θ 50 [CAD] θ 50 [CAD]

0
∂ θ 50 / ∂p in [CAD/bar]

0.6
−2

||S c (z)||2
0.5
−4
0.4
−6
0.3
5 10 15 5 10 15
θ 50 [CAD] θ 50 [CAD]

PRF100, LL PRF100, HL PRF0, LL PRF0, HL

Figure 6.14 The closed-loop noise sensitivity depends on θ50 partial derivatives
with respect to Tin , p in , Q ctot and the S c norm. This figure presents these quantities
as a function of θ50 for PRF0, PRF100 with p IMEPg = 5 bar (LL) and 15 bar (HL).

Partial derivatives of θ50 and ||S c ||2 with constant controller gains are presen-
ted in Fig. 6.14 for p IMEPg = 5 and 15 bar. The partial-derivate magnitudes were
clearly larger at low p IMEPg and late θ50 . It can also be seen that PRF0 was more
sensitive to Q ctot and p in whilst PRF100 was more sensitive to Tin . The closed-loop
noise sensitivity ||S c ||2 was higher for PRF100 and decreased with θ50 . Similar ex-
perimental PRF trends were presented in [Sjöberg and Dec, 2005], where Tin was
adjusted according to PRF.
The standard deviation σθSOI , computed using (6.18) is presented in Fig. 6.15.
It can be seen that σSOI decreased with p IMEPg and θ50 . Overall, σSOI was also
higher for PRF0 which agrees with the observed trends in Figs. 6.8 and 6.13.

120
 6.5 Conclusions

Noise Sensitivity
0.3
PRF100, HL
PRF100, LL
0.25 PRF0, HL
PRF0, LL
σθSOI [CAD]

0.2

0.15

0.1

0.05
4 6 8 10 12 14
θ 50 [CAD]

Figure 6.15 Injection-timing standard deviation σSOI as a function of θ50 , com-

puted using (6.18). The standard deviation was altogether higher for PRF0 which
agrees with the observed trends in Figs. 6.8 and 6.13.

Rules of Thumb
To summarize, the simulation results provided the following rules of thumb for
PI-controller tuning:

1. Robustness was limited by operating points with large ∂θ50 /∂θSOI . These
were found at late θ50 and low p IMEPg . At these points, ∂θ50 /∂θSOI incre-
ased with PRF value, meaning that robustness constraints became more
restrictive for higher PRF values.

2. Noise sensitivity was higher at low loads and late combustion timings
where gains from disturbances to θ50 were higher. The noise sensitivity was
also higher for lower PRF values due to an increased Q ctot and p in sensitivity.

3. A robust and noise-insensitive controller-gain choice for all fuels and ope-
rating points is given by 0.2 < k p < k I < 0.35.

6.5 Conclusions
This chapter showed that the θSOI interval for which θ50 is controllable is limited
between a low-gain limit for early θSOI and a misfire limit for late θSOI . In order
to obtain early combustion timings for higher PRFs, Tin had to be adjusted as a

121
Chapter 6. Proportional-Integral Combustion-Timing Control

function of PRF. Even after Tin adjustment, the PRFs have different θ50 characte-
ristics, partly due to the varying NTC behavior for the different fuels. The com-
putation of suitable θSOI limits is a part of the controller design where such limits
could be a function of fuel, load and intake conditions.
Optimal PI-controller gains found through simulation were limited by the
high system gain at late θ50 and low p IMEPg . Controller measurement-noise sen-
sitivity was also found to be higher at this operating point due to an increased θ50
sensitivity to load and intake-condition variation. Controller gain requirements
also varied with the PRF value: robustness was lower for higher PRFs due to an
increased system gain, and noise sensitivity was higher for lower PRFs due to a
higher Q ctot and p in sensitivity.

122
7
Model-Based Control of
Combustion Timing and
Ignition Delay

7.1 Introduction
A sufficiently long ignition delay τ is a prerequisite for the fuel / air mixing le-
ading to premixed low-temperature combustion, as described in Sec. 1.3. Mea-
nwhile, it is also important that the combustion timing θ50 (4.3) is sufficiently
well timed for high thermodynamic efficiency. A too early θ50 leads to increased
heat-transfer and inefficient pressure build-up during the compression stroke,
and a late θ50 results in high exhaust losses and a lowered combustion efficiency.
Simultaneous control of τ and θ50 is an important component for a successful
implementation of partially premixed combustion, and when controlling these
two variables, one has to pay attention to their coupling through θSOI and ther-
modynamic in-cylinder conditions.
This chapter studies model-based multi-cylinder control of τ and θ50 with
combined actuation of the gas-exchange and fuel-injection systems. The objec-
tive is to regulate τ and θ50 during load disturbances and set-point changes. This
is an under-determined control problem due to more output than input variables
for the engine setup used.
A model predictive control (MPC) is suggested, see Sec. 3.1. This is a suita-
ble design choice for multiple input/output systems with actuator constraints.
The controller obtains τ and θ50 from in-cylinder pressure measurement and
heat-release analysis, and the ignition-delay model M1 (see Sec. 2.4) is linearized
every engine cycle for model-based prediction. The MPC feedback loop is illus-
trated in Fig. 7.1, where the multiple cylinders are indicated by the (bold type)
vector notation. The MPC computes fuel-injection timings and valve positions
of the two EGR paths and the fast thermal-management (FTM) system. Valve po-

123
Chapter 7. Model-Based Control of Combustion Timing and Ignition Delay

Disturbance

τr , θ50
r
θ SOI τ, θ 50
MPC Engine
θ EGR
θ FTM

Linearization

−1

Figure 7.1 This chapter presents and studies the following MPC feedback loop.
The controller obtains τ and θ 50 from in-cylinder pressure measurement and
heat-release analysis, and then linearizes the ignition-delay model M1, presented
in Sec. 2.4. The controller computes fuel-injection timings θ SOI and valve posi-
tions of the two EGR paths θ EGR and the fast thermal-management system θ FTM .
Multiple cylinders are indicated by the (bold type) vector notation.

sitions determine gas-mixture temperature and exhaust-gas recirculation (EGR)

ratio, see Fig. 2.3.
Previous research on this subject was presented in [Lewander et al., 2008],
where MPC was used to control τ for premixed operation whilst θ50 was kept wi-
thin an acceptable range by injection-timing adjustment. Karlsson et al. [2008]
controlled τ and θ50 in conventional diesel combustion using a linear-quadratic
regulator (LQR) to minimize emissions using the back-pressure valve and θSOI
as system inputs. The main contributions of the work presented in this chap-
ter are the use of a physics-based τ model for MPC, and the use of additional
gas-exchange actuators.
The modeling and linearization procedures are introduced in Sec. 7.2, and
the controller design is presented in Sec. 7.3. The main part of this chapter covers
an experimental controller evaluation, which is presented in Sec. 7.4. Discussion
and conclusions are given in Sec. 7.5.

7.2 Modeling
Model M1 (see, Sec. 2.4) was used to model τ, whereas a calibrated stat-
ic model was used to model the gains from gas-exchange valve positions to
intake-manifold composition and temperature. This section describes these
models and how they were used in the controller design.

124
 7.2 Modeling

Ignition Delay Model

The ignition-delay model M1, without p dependence showed comparable per-
formance to more detailed models but had a lower computational complexity,
see Sec. 2.4. The model is given by
α
τ = A[O2 ] e E a /R̃T (7.1)

where E a is an apparent activation energy, R̃ is the universal gas constant and T

and [O2 ] are the mean cylinder temperature and oxygen concentration between
θSOI and θ10 . Here, the in-cylinder temperature was computed using the adia-
batic compression relation
³ V ´γ−1
IVC
T = TIVC (7.2)
V
where γ was held constant. The oxygen concentration was then given by
[O2 ]IVC
[O2 ] = VIVC (7.3)
V
where [O2 ]IVC is the in-cylinder oxygen concentration at intake-valve closing.
It was computed using (2.41) and an estimated EGR mass flow, as described in
Sec. 2.3. The following expression for τ
µ ¶µ Z θ +τ ¶α
Ea 1 SOI [O2 ]IVC
τ = Aexp Z θSOI +τ µ ¶γ−1 VIVC d θ (7.4)
R̃ VIVC τ θSOI V (θ)
TIVC dθ
τ θSOI V (θ)
is obtained when substituting for (7.2) and (7.3), with assumed constant Nspeed .
Model parameters A, E a and α were found using the identification procedure
presented in Sec. 2.4 at p IMEPg = 5 bar and Nspeed = 1200 rpm.

Gas-Exchange System Model

Simple static models, determined from experimental data, were used to relate
changes in gas-system valve positions to changes in TIVC and [O2 ]IVC . In Fig. 7.2,
TIVC and [O2 ]IVC are displayed as functions of the high and low-pressure EGR val-
ve positions θLP , θHP , and the hot-path valve positions θhot . The cool-path valve
position was changed by setting

θcool = cos−1 (1 − cos(θhot )) (7.5)

in order to keep an approximately constant total valve-opening area. This app-

roach was previously used in [Widd et al., 2009].
This modeling approach is of course an oversimplification. One argument for
using this model instead of the dynamic gas-exchange model in Sec. 2.3, besides
reducing computational complexity, was that the valve-actuator dynamics were
slower than the pressure dynamics.

125
Chapter 7. Model-Based Control of Combustion Timing and Ignition Delay

Static FTM-Valve Effect on Inlet-Manifold Temperature

330
p IMEP = 5 bar
p IMEP = 10 bar
320
Tin [K]

310

300

290
10 20 30 40 50 60 70 80 90
θ hot [deg]

Static EGR-Valve Effect on Oxygen Concentration

12
θ HP , p IMEP = 5 bar
θ HP , p IMEP = 10 bar
[O2 ]IVC [mol/m3 ]

θ LP , p IMEP = 5 bar
10 θ LP , p IMEP = 10 bar

0 10 20 30 40 50 60 70 80 90
θ HP,LP [deg]

Figure 7.2 TIVC and [O2 ]IVC as functions of θLP , θHP and θhot at two different
loads and Nspeed = 1200 rpm. These experimentally obtained functions were used
to model the relation between the gas-exchange valve positions and the intake
manifold conditions. Note that the gain from θHP decreased with load. This was
because of the decreased pressure difference over the exhaust and intake mani-
folds due to increased turbocharger boost.

Differentiation and Linearization

The presented models are not directly applicable for linear MPC. Model linear-
ization was therefore necessary to obtain linear equality constraints. The app-
roach taken was to linearize (7.4) and the trends in Fig. 7.2 numerically every
engine cycle to provide a good model approximation. Since (7.4) is an implicit
relation in τ, the partial derivatives of τ with respect to θSOI , TIVC and [O2 ]IVC

126
 7.2 Modeling

were approximated by keeping τ constant in the integration limit in (7.4)

∂τ τ(θSOI + ∆θSOI /2) − τ(θSOI − ∆θSOI /2)

≈
∂θSOI ∆θSOI

∂τ τ(TIVC + ∆TIVC /2) − τ(TIVC − ∆TIVC /2)

≈ (7.6)
∂TIVC ∆TIVC

∂τ τ([O2 ]IVC + ∆[O2 ]IVC /2) − τ([O2 ]IVC − ∆[O2 ]IVC /2)
≈
∂[O2 ]IVC ∆[O2 ]IVC

This approximation is motivated by the fact that small increments in T and

[O2 ] do not change τ significantly. The partial derivatives in (7.6) make it pos-
sible to relate changes in ∆TIVC , ∆θSOI , ∆[O2 ]IVC to changes in τ and θ50 on a
cycle-to-cycle basis

 
µ ¶ ∆θSOIi (k)
∂τi ∂τi ∂τi  
τi (k + 1) = τi (k) + i  ∆TIVC (k) 
∂θSOI ∂TIVC ∂[O2 ]IVC
∆[O2 ]IVC (k)
θ50i (k + 1) = θ50i (k) + ∆θSOIi (k) (7.7)
 
µ ¶ ∆θSOIi (k)
∂θ ∂τi ∂τi ∂τi  
+ i  ∆TIVC (k) 
∂t ∂θSOI ∂TIVC ∂[O2 ]IVC
∆[O2 ]IVC (k)

where k is the cycle index, ∆ is the forward-difference operator, i is the cylinder

index, and d θ/d t is crank angles per millisecond at the current Nspeed , needed
here since τ and θ50 have different units.
The partial derivatives ∂TIVC /∂θhot , ∂[O2 ]IVC /∂θLP and ∂[O2 ]IVC /∂θHP , comp-
uted from the slopes in Fig. 7.2, relate the gas-system valve positions to the sys-
tem outputs τ and θ50 accordingly

 
∆θSOIi (k)
µ ¶ 
∂τi ∂τi ∂τi ∂τi  ∆θhot (k) 
τi (k + 1) = τi (k) +  
i
∂θSOI ∂θhot ∂θHP ∂θLP   ∆θHP (k) 


∆θLP (k)
  (7.8)
∆θSOIi (k)
µ ¶ 
∂θ50i ∂θ50i ∂θ50i ∂θ50i  ∆θhot (k) 
θ50i (k + 1) = θ50i (k) +  
∂θSOIi ∂θhot ∂θHP ∂θLP  ∆θHP (k) 


∆θLP (k)

127
Chapter 7. Model-Based Control of Combustion Timing and Ignition Delay

The complete linear state-space model can be written on more compact form
 
∆θ SOI (k)
à ! à !  
θ 50 (k + 1) θ 50 (k)  ∆θhot (k) 
= +B 
 ∆θ (k) 
 (7.9)
τ(k + 1) τ(k)  HP 
∆θLP (k)

where
à !
θ 50 (k) ¡ ¢T
= θ501 (k) ... θ506 (k) τ1 (k) ... τ6 (k)
τ(k)
¡ ¢T
∆θ SOI (k) = ∆θSOI,1 (k) ... ∆θSOI,6 (k)
 ∂θ ∂θ50,1 ∂θ50,1 ∂θ50,1 
50,1
··· 0
 ∂θSOI,1 ∂θhot ∂θHP ∂θLP 
 
 .. .. .. .. .. .. 
 . . . . . . 
 
  (7.10)
 ∂θ50,6 ∂θ50,6 ∂θ50,6 ∂θ50,6 
 0 ··· 
 ∂θSOI,6 ∂θhot ∂θHP ∂θLP 
 
B = 
 ∂τ1 ∂τ1 ∂τ1 ∂τ1 
 ... 0 
 ∂θ ∂θhot ∂θHP ∂θLP 
 SOI,1 
 .. .. .. .. .. .. 
 
 . . . . . . 
 
 ∂τ6 ∂τ6 ∂τ6 ∂τ6 
0 ···
∂θSOI,6 ∂θhot ∂θHP ∂θLP

The system is not controllable due to more outputs than inputs, which is indica-
ted by the fact that the controllability matrix C
¡ ¢
C= B ... B (7.11)

has less than 12 independent columns. The system outputs τ and θ 50 can there-
fore not be controlled independently and we are left with two options. Combus-
tion timings θ 50 can either be controlled cylinder individually by varying θ SOI
and then letting gas-exchange actuators control τ. The other option is to control
τ cylinder individually and let the gas-exchange actuators control θ 50 . The first of
the two approaches was taken here. The reason for this choice was the nonlinear
relationship between θ SOI and τ (see Fig. 6.4), and the monotone relationship
between τ and TIVC , [O2 ]IVC . It was also shown in Sec. 2.4 that it is difficult to ac-
curately model the sign of ∂τ/∂θSOI when θSOI is close to TDC. Such model errors
could lead to unwanted closed-loop behavior if θ SOI is set to control τ.
The desired controller can be obtained by MPC-weight tuning, which is cove-
red in the following sections.

128
 7.3 Model Predictive Control Formulation

7.3 Model Predictive Control Formulation

The MPC problem was formulated accordingly

Hp ³
X r
minimize ω1 ||θ50 (k) − θ 50 (k)||22 + ω2 ||τr (k) − τ(k)||22
∆θ SOI , ∆θHP , k=1 ´
∆θLP , ∆θhot + ω3 θHP (k)2 + ω4 θLP (k)2
X ³
+ ω5 ||∆θ SOI (k)||22 + ω6 ∆θhot (k)2
k∈k Hc ´
+ ω7 ∆θHP (k)2 + ω8 ∆θLP (k)2 (7.12)

 
θ SOI
 θhot 
subject to θ l ≤  
 θHP  ≤ θ
u

θLP

and. (7.9) for k = 0, . . . , H p − 1

Here, k is the cycle index, and || · ||2 is the Euclidian norm in R6 . Initial condi-
tions at the current cycle k = 0 are obtained from measurements. The first sum
penalizes θ50 and τ set-point error, and the usage of EGR over the prediction ho-
r
rizon H p . Set points θ50 and τr are considered to be precomputed as a function
of the engine operating point. It was decided to penalize EGR-valve opening area
to favor flow over the turbine and to not use more EGR than needed. The terms
in the second sum penalize control action over the control horizon Hc . The cost
function should be minimized subject to absolute constraints on actuators and
the linearized system dynamics in (7.9).

MPC Design and Implementation

In order for the controller to overlook the slow valve actuators, H p was set equal
to 50 engine cycles. Moreover, the inputs were allowed to change nonequidistan-
tly at samples k ∈ k H c over the control horizon to decrease the number of varia-
bles in the optimization problem, and allow for shorter computation times, see
Fig. 7.3.
The relation between the weights ω1−2 and ω5−8 determines the trade-off bet-
ween tracking performance and controller sensitivity to cycle-to-cycle variation
and model error. The weights ω6−8 determine how fast the valve positions are
allowed to change, and were chosen to conform with physical limitations. The
choice of ω3−4 is a trade-off between the ability to supply EGR to increase τ and
gas-exchange efficiency. Input bounds for EGR-valve positions were chosen so
that the controller would operate in intervals where the [O2 ]IVC -gains were non-

129
Chapter 7. Model-Based Control of Combustion Timing and Ignition Delay

Table 7.1 Weights and constraints for (7.12).

ω1 ω2 ω3 ω4
2 4
10 10 5000 3000
ω5 ω6 ω7 ω8
−2
0.1 10 400 800
3 ≤ θHP ≤ 40 20 ≤ θLP ≤ 50 5 ≤ θhot ≤ 85 −40 ≤ θSOI ≤ 10

zero, see Fig. 7.2. Other actuator bounds were chosen to fulfill physical limita-
tions, and θSOI was limited to avoid misfire, see Fig 6.2.
The weights and constraints used are presented in Table 7.1. Solution trajec-
tories with these weights are presented in Fig. 7.3 for an arbitrary initial condi-
tion. The dashed black lines in the upper part of Fig. 7.3 are set points τr and
r
θ50 . The cylinder numbers are indicated by different colors, cylinder 1 being the
upper cylinder in Fig. 2.3.
The weights were tuned to obtain the desired behavior where θ SOI controls
cylinder-individual θ 50 . This was done by prioritizing θ 50 tracking, and let the
slow gas-system actuators control the mean τ. Note that the grid over which θ SOI
is allowed to change is denser initially than for the gas system valves. This is be-
cause the actuation of θ SOI is much faster than for the valve positions.
Solving (7.12) is a quadratic program (QP) and was solved in LabVIEW using
the QP-solver VI. The solver used the previous-cycle solution and active set as
initial guesses for the next cycle to shorten computation times. Early termina-
tion was also used for the solver to finish within one engine cycle. These are well
known methods for speeding up MPC execution, see [Wang and Boyd, 2010]. The
average computation time for differentiating Eqs. (7.6) and constructing the QP
matrices was 10 ms, while it took 25 ms on average to solve (7.12). These com-
putations were repeated every engine cycle after sampling of the previous-cycle
cylinder pressure.

7.4 Experimental Results

Controller experiments were conducted in steady-state and during load and
r
speed changes. In steady state, the objective was to track step changes in θ50 and
τr , whereas the objective during load and speed changes was to regulate θ 50 and
τ at constant set points. The engine load p IMEPn was changed by varying the inj-
ection durations, whilst the common-rail pressure p rail was held constant at 800
bar. Both p IMEPn and p rail were controlled using PI controllers with feedforward.
Engine-speed ramps were performed by manually changing the engine-brake
speed set point.

130
 7.4 Experimental Results

Cyl 1 Cyl 2 Cyl 3 Cyl 4 Cyl 5 Cyl 6

2.6 10

θ 50 [CAD]
τ [ms]

2.4
6

2.2 4

2
20 40 20 40
cycle [-] cycle [-]

18 −11

16 −12

θ SOI [CAD]
θ HP [deg]

14 −13

12 −14

10 −15
20 40 20 40
cycle [-] cycle [-]

80 θ cool θ hot 32
θ hot /θ cool [deg]

70 30
θ LP [deg]

60 28

50 26

40 24
20 40 20 40
cycle [-] cycle [-]

Figure 7.3 Solution trajectories of (7.12) using the weights and constraints in
Table 7.1 and an arbitrary initial condition. The inputs were allowed to change
at predefined cycle indices over Hc . The weights were chosen to prioritize θ 50 tra-
cking, and let the slower gas-system actuators control the mean τ. Note that the
samples for which θ SOI was allowed to change was denser initially than for the
gas-system valves. This was because control action of θ SOI was much faster than
for the valve positions.

Set-Point Tracking
Set-point tracking performance was evaluated by keeping p IMEPn and Nspeed
r
constant and varying θ50 and τr . System inputs and outputs during 800 cycles
r
of θ50 step changes are presented in Fig. 7.4. The controller weights ωi (see Table

131
Chapter 7. Model-Based Control of Combustion Timing and Ignition Delay

3) were set so that tracking of θ 50 was done by changing θ SOI , τ was then jointly
controlled by the gas-exchange system actuators. This tuning resulted in regula-
tion of the mean cylinder τ.
As θ 50 was delayed at cycle 950 in Fig. 7.4, τ decreased due to the increased
temperature at θ SOI . This forced θhot to close while θHP,LP opened. The control-
ler tried to find the lowest possible θHP in stationarity for higher turbine mass
flows. Some chattering is visible in the gas-exchange actuators. This was due to
stochastic cycle-to-cycle variation in θ 50 and τ which could have been reduced
by introducing filtering or increasing ω6−8 .
A zoom-in around cycle 950 is presented in Fig. 7.5. Here, it can be seen that
θ 50 reached the new set point in 5 cycles. There was an internal delay of 5 cycles
from set point to the controller, caused by the communication from the user in-
terface and the real-time system. The gas system managed to adjust for the τ de-
crease in 50 cycles where cylinder-to-cylinder variation created a τ-distribution
among the cylinders. This variation could be caused by non-uniform EGR distri-
bution to the different cylinders or different cylinder-wall temperatures. It can
be seen that τ6 was consistently shorter whilst τ1 was the longest. A hypothetical
explanation for this is that cylinder 1 is closer to the high-pressure EGR path, see
Fig. 2.3.
In Fig. 7.6, system inputs and outputs are displayed during 1400 cycles for
which step changes in τr were made. The tracking of τr was realized by varying
θhot and θHP,LP whilst θ SOI was varied to keep θ 50 constant. The high-pressure
EGR valve opened too much initially which gave a slight overshoot in τ. A
zoom-in around cycle 550 is presented in Fig. 7.7. Here, it can be seen that θ SOI
was varied to keep θ 50 within 0.6 CAD whilst τ reached the new set point in 50
cycles.

Load Changes
In Fig. 7.8, system inputs and outputs are displayed during 1000 cycles for which
r
p IMEPn steps were made between 6 and 10 bar, and p IMEPn reached its new set
point in 20 cycles. The ignition delays decreased as p IMEPn was increased due to
increased cylinder temperature and richer cylinder mixtures. This forced θHP,LP
to increase. The FTM system was limited to the cold-flow upper limit during the
higher p IMEPn values. Adjustments in θ SOI managed to keep θ 50 within 1 CAD.
In-cylinder data at p IMEPn = 6 (dashed) and p IMEPn = 10 bar (solid) are pre-
sented in Fig. 7.9. The figure shows pressure and heat-release rates for the diffe-
rent cylinders, and the injection current from cylinder 1. The controller manages
to maintain constant θ 50 and τ despite the p IMEPn difference.

Speed Changes
In Fig. 7.10, system inputs and outputs are displayed during 2000 cycles for
which the engine speed Nspeed was varied between 1200 and 1500 rpm. The igni-

132
 7.4 Experimental Results

Cyl 1 Cyl 2 Cyl 3 Cyl 4 Cyl 5 Cyl 6

4
10

θ 50 [CAD]
3.5 8
τ [ms]

3 6

4
2.5
800 1,000 1,200 1,400 800 1,000 1,200 1,400
cycle [-] cycle [-]

−10
60

θ SOI [CAD]
θ LP [deg]

−15
40
−20

20
−25
800 1,000 1,200 1,400 800 1,000 1,200 1,400
cycle [-] cycle [-]

100
40
θ cool /θ hot [deg]

80
θ HP [deg]

60
40 20

20
0 0
800 1,000 1,200 1,400 800 1,000 1,200 1,400
cycle [-] cycle [-]

Figure 7.4 System inputs and outputs during step changes in θ50 r .θ
hot is indica-
ted in red and θcool in blue. The black dashed lines in the upper diagrams are θ50 r
r
and τ , whilst the dash-dotted lines are constraints on θLP and θHP .

tion delays decreased as Nspeed was increased. Probably due to increased engine
temperatures and cylinder-gas turbulence levels [Heywood, 1988]. This forced
θHP,LP and θcool to open. The FTM system was limited to the cold-flow limit at
r
Nspeed = 1500 rpm, and θ SOI managed to keep θ 50 within 1 CAD from θ50 .

133
Chapter 7. Model-Based Control of Combustion Timing and Ignition Delay

Cyl 1 Cyl 2 Cyl 3 Cyl 4 Cyl 5 Cyl 6

4
10

θ 50 [CAD]
3.5 8
τ [ms]

6
3
4

2.5 2
920 940 960 980 1,000 920 940 960 980 1,000
cycle [-] cycle [-]

−12
60
−14

θ SOI [CAD]
θ LP [deg]

−16
40
−18
−20
20
−22
920 940 960 980 1,000 920 940 960 980 1,000
cycle [-] cycle [-]

100
40
θ cool /θ hot [deg]

80
θ HP [deg]

60
40 20

20
0 0
920 940 960 980 1,000 920 940 960 980 1,000
cycle [-] cycle [-]

Figure 7.5 A zoom in around cycle 960 in Fig. 7.4 where θ 50 reached a new set
point in 5 cycles. There is a visible internal delay from the set point to the control-
ler of about 5 cycles. The gas system managed to adjust for the τ decrease in 50
cycles.

7.5 Discussion and Conclusions

This chapter presented a physics-based MPC for control of θ50 and τ in
a multi-cylinder engine. The controller was tuned for fast control of θ50
compared to τ, which resulted in a controller behavior where θSOI control-
led cylinder-individual θ50 and the cylinder-averaged τ was controlled by
gas-exchange system actuation. The motivation for this design was the simp-

134
 7.5 Discussion and Conclusions

Cyl 1 Cyl 2 Cyl 3 Cyl 4 Cyl 5 Cyl 6

4 8

θ 50 [CAD]
3.5
τ [ms]

6
3
5

2.5 4
500 1,000 1,500 500 1,000 1,500
cycle [-] cycle [-]

−14
60
−16

θ SOI [CAD]
θ LP [deg]

40 −18

−20
20
−22
500 1,000 1,500 500 1,000 1,500
cycle [-] cycle [-]

100
40
θ cool /θ hot [deg]

80
θ HP [deg]

60
40 20

20
0 0
500 1,000 1,500 500 1,000 1,500
cycle [-] cycle [-]

Figure 7.6 System inputs and outputs during 1400 cycles for which steps in τr
were made. The tracking of τr was realized by varying θhot and θHP,LP whilst θ SOI
kept θ 50 constant.

ler relation between intake conditions and τ, compared to the relation between
θSOI and τ. Additional actuation techniques such as variable valve timings, varia-
ble compression ratio or spark ignition could be applied to control both θ50 and
τ cylinder individually
The suggested controller was successful in tracking τ and θ50 set points with
response times of 50 and 5 cycles, respectively. Both in stationarity and du-
ring load and speed changes. A general observation was that MPC-weight tun-
ing was a trade-off between short response times and closed-loop robustness

135
Chapter 7. Model-Based Control of Combustion Timing and Ignition Delay

Cyl 1 Cyl 2 Cyl 3 Cyl 4 Cyl 5 Cyl 6

4 8

θ 50 [CAD]
3.5
τ [ms]

6
3
5

2.5 4
450 500 550 600 650 450 500 550 600 650
cycle [-] cycle [-]

60 −16

θ SOI [CAD]
−18
θ LR [deg]

40 −20
−22
20 −24
450 500 550 600 650 450 500 550 600 650
cycle [-] cycle [-]

100
40
θ cool /θ hot [deg]

80
θ HP [deg]

60
40 20

20
0 0
450 500 550 600 650 450 500 550 600 650
cycle [-] cycle [-]

Figure 7.7 A zoom in around cycle 550 in Fig. 7.6. The ignition delay reached its
new set point after 50 cycles and θ SOI kept θ 50 within 0.6 CAD during the step
change.

to model-errors and cycle-to-cycle variation. Control action would increase if

the control-action weights were to be decreased. This could however lead to
system-output overshoot during set-point changes and increased actuator chat-
tering in steady state, partly as result of insufficient model accuracy. As a result,
the controller was tuned to give slow but robust performance. The gas-system
model used in this work was limited to the static relation between valve positions
and the intake manifold gas state in a small operating range. It is, however, possi-
ble that a dynamic model of valve actuators and the gas-system dynamics could

136
 7.5 Discussion and Conclusions

10
3.5

θ 50 [CAD]
8
τ [ms]
3
6

2.5
4
200 400 600 800 1,000 200 400 600 800 1,000
cycle [-] cycle [-]

−10
60
−12

θ SOI [CAD]
θ LP [deg]

−14
40
−16
−18
20
−20
200 400 600 800 1,000 200 400 600 800 1,000
cycle [-] cycle [-]

100
40
θ cool /θ hot [deg]

80
θ HP [deg]

60
40 20

20
0 0
200 400 600 800 1,000 200 400 600 800 1,000
cycle [-] cycle [-]

12 1.8
p IMEPn [bar]

1.6
θ DOI [deg]

10
1.4
8
1.2
6
1
200 400 600 800 1,000 200 400 600 800 1,000
cycle [-] cycle [-]

Figure 7.8 System inputs and outputs during 1000 cycles for which steps in
r
p IMEPn were made. In addition to the signals displayed in the previous figures,
p IMEPn is plotted together with its set point in the lower left figure. Injection du-
rations θDOI are presented in the lower-right subdiagram. In-cylinder data from
this experiment are presented in Fig. 7.9.

137
Chapter 7. Model-Based Control of Combustion Timing and Ignition Delay

120

100

80
p [bar]

60

40

20

0
−20 −15 −10 −5 0 5 10 15 20 25
θ [CAD]

Figure 7.9 In-cylinder data from the experiment in Fig. 7.8 at p IMEPn = 6 (da-
shed) and p IMEPn = 10 bar (solid). The figure shows pressures and heat-release
rates for the different cylinders and the injection current from cylinder 1. The con-
troller maintains constant θ 50 and τ despite the p IMEPn difference.

improve performance. This would also allow for the gas-exchange efficiency to
be included in the MPC cost function.
The MPC framework provided a simple way of prioritizing system output be-
havior. It also took interaction effects into account. Input constraints and the
cost of using EGR were also incorporated. Comparable controller performance
could probably be obtained by adopting an approach where θ50 is controlled with
θSOI and cylinder-individual controllers, and then let the mean τ be controlled
by the gas-system valve positions, see [Karlsson et al., 2008]. This approach de-
mands less on-line computations but is less general compared to the framework
presented here, where input and output constraints can be incorporated. The
MPC could also be extended to cover p IMEPn control by adding the injected fuel
amount and its effect on τ to the model. Variation in fuel amount could in this
way also be included in the prediction to reduce set-point deviation during load
transients.
A sufficient τ was here considered to be an indicator for low temperature
combustion with favorable emission properties. An assumption that was shown
to be accurate in [Karlsson et al., 2008; Lewander et al., 2008]. In future work, this
controller could be evaluated with emission measurements to conclude if this
hypothesis holds or if supplementary control actions need to be taken.

138
 7.5 Discussion and Conclusions

Cyl 1 Cyl 2 Cyl 3 Cyl 4 Cyl 5 Cyl 6

3.6 10
3.4 9

θ 50 [CAD]
τ [ms]

3.2
8
3
2.8 7

2.6 6
0 500 1,000 1,500 0 500 1,000 1,500
cycle cycle

−14
60
θ HP / θ LP [deg]

−16

θ SOI [CAD]
40 −18
−20
20
−22
0 −24
0 500 1,000 1,500 0 500 1,000 1,500
cycle cycle

100 1,600
θ cool /θ hot [deg]

80
Nspeed [rpm]

60 1,400
40
20 1,200
0
0 500 1,000 1,500 0 500 1,000 1,500
cycle cycle

Figure 7.10 System inputs and outputs during 2000 cycles for which Nspeed was
varied, see the lower-right subdiagram. Both EGR-valve positions are now plotted
together in the mid-left figure.

139
8
Pilot Injection

8.1 Introduction
It was discovered in [Manente et al., 2009] that long ignition delays in
single-injection PPC give rise to high pressure-rises rate due to violent HCCI-like
combustion. A long ignition delay creates in-cylinder mixtures where the major-
ity of the injected fuel reaches high-temperature reactions simultaneously, which
result in high heat-release and pressure-rise rates. An example of this can be seen
at the higher-load operating point in Fig. 7.9. A high pressure-rise rate is an in-
dicator for high audible noise levels and could also lead to mechanical engine
damage. The pressure-rise rate therefore has to be kept below certain levels in or-
der to ensure silent and safe engine operation. Previous research by Tsurushima
et al. [2009] implies that pressure oscillations resulting from violent combus-
tion rates are able to break insulating gas boundary layers in the cylinder. High
pressure-rise rates could therefore result in increased heat-transfer flux to the
cylinder walls. The issue of having high pressure-rise rates is not as severe in
conventional diesel engines where the heat-release rate is limited by the rates of
fuel-injection and fuel-air mixing.
A means of counteracting the pressure-rise rate problem is to introduce a pil-
ot injection, e.g., by having a smaller fuel injection earlier during the compres-
sion stroke and then inject the majority of the fuel amount closer to TDC in a
main injection. Optical OH chemiluminescence experiments and computatio-
nal fluid dynamics (CFD) simulations suggest that the pilot injection provides a
background mixture whose reactions increase in-cylinder temperature and as-
sist autoignition of the the main injection [Tanov et al., 2014; Solsjö, 2014]. The
reduced ignition delay of the main injection increases in-cylinder stratification
during combustion with decreased combustion rates as a result.
In-cylinder data showing the effect of introducing a pilot injection is pre-
sented in Fig. 8.1. The two fuel-injection configurations produce the same load,
p IMEP = 5 bar, and the same combustion timing θ50 = 6 CAD, but with different
pressure-rise rates d p max = 11 bar/CAD (blue) and d p max = 30 bar/CAD. Pilot in-
jections are also used in conventional diesel engines, both to improve low-load

140
 8.1 Introduction

120

100

80

60
p [bar]

40

20

−20
−25 −20 −15 −10 −5 0 5 10 15 20 25
θ [CAD]

Figure 8.1 In-cylinder data showing the effect of introducing a pilot injection.
The two fuel-injection configurations produce the same load, p IMEP = 5 bar, and
the same combustion timing θ50 = 6 CAD, with different maximum pressure-rise
rates d p max = 11 bar/CAD (blue) and d p max = 30 bar/CAD (red). Pressure-rise
rates are indicated by the dashed tangent lines.

performance [MacMillan et al., 2009; Osuka et al., 1994] and to decrease emis-
sions and engine-noise levels [Kiencke and Nielsen, 2000].
With additional injections, the amount of calibration work for optimized
engine performance at different operating points grows exponentially [Meyer,
2011]. It would therefore be advantageous to have a fuel-injection controller
that automatically adjusts injection timings and the fuel distribution among
the injections depending on the engine operating point. Previous work on
pilot-injection combustion control in low-temperature combustion concepts
has been investigated in [Ott et al., 2013; Eichmeier et al., 2012; Ekholm et al.,
2008; Kokjohn et al., 2009]. Whereas previous works have focused either on ca-
libration and control in open-loop, or on decentralized PI control of θ50 and
d p max , this chapter introduces an MPC that aims to decouple control of θ50 and
d p max , where control of d p max is formulated as an output constraint. The sug-
gested feedback loop is presented in Fig. 8.2, where a Kalman filter is used to
attenuate stochastic output variation.
This chapter begins with a presentation of experimental engine-performance
characteristics in terms of efficiency, emissions and maximum pressure-rise rate
controllability, with respect to pilot-injection parameters in the low-to-mid load

141
Chapter 8. Pilot Injection

Disturbance, e

c r m e e
d p max , θ50 θSOI d p max , θ50
MPC Engine
rp

d p max , θ50
Kalman filter

Figure 8.2 This chapter studies the following feedback loop where the ratio of
pilot-fuel injected r p and main-injection timing θSOIm are adjusted to keep the
c
pressure-rise rate d p max below an upper limit d p max , while the combustion ti-
r . A Kalman filter was used to attenuate sto-
ming θ50 should follow a set point θ50
chastic output variation.

range of the engine. The experimental results are then used to design the con-
troller in Fig. 8.2. An experimental controller evaluation during engine load and
speed changes is also presented in this chapter.

Controlled Outputs
The maximum pressure rise rate d p max is here defined as

dp
d p max = max (8.1)
θ dθ
Due to the high cycle-to-cycle variation in d p max , filtering is necessary for d p max
e
to be used as a feedback variable. The computed d p max was therefore modeled
as the cycle mean d p max , corrupted with additive Gaussian noise e with standard
deviation σd p max
e
d p max = d p max + e d p max , e d p max ∼ N (0, σ2d p max ) (8.2)

Stochastic variation in θ50 was also modeled as additive Gaussian noise accor-
ding to
e
θ50 = θ50 + e θ50 , e θ50 ∼ N (0, σ2θ50 ) (8.3)
e e
A Kalman filter was then used to recover d p max and θ50 from d p max and θ50 . The
Kalman filter will be presented in Sec. 8.3.

Input Variables
The input variables considered are the fuel-injection timings and durations defi-
ned by the current pulses sent to the injector, see Fig. 8.3. The injector-current ri-
m p
sing flanks indicate the start of the main and pilot injections θSOI , θSOI , that occur

142
 8.1 Introduction

Input Variables
2

1.5
Injector Current [a.u]

p
1 θ SOI θm
SOI

0.5

0
p
θh θ
DOI
θh θm
DOI
−0.5

−50 −40 −30 −20 −10 0

θ [CAD]

Figure 8.3 Injector-current signal with a pilot (p) and a main (m) injection toge-
x , θ x , and θ .
ther with definitions of θSOI DOI h

m p
after an hydraulic injector delay, θh . The fuel-injection durations θDOI and θDOI
are defined as the difference between the injector-current pulse width and θh .
The delay therefore determines the minimum current-pulse duration for which
fuel is injected into the cylinder. It was here assumed to be constant at 0.25 ms.
m p m p
Instead of studying the effects of θSOI , θSOI , θDOI and θDOI explicitly, the pilot
ratio r p
p
θDOI
rp = p m
(8.4)
θDOI + θDOI
the injection separation d SOI
m p
d SOI = θSOI − θSOI (8.5)
m
the main-injection timing θSOI , and the total injection duration
tot m p
θDOI = θDOI + θDOI (8.6)

were considered.
tot m
These variables were chosen because θDOI and θSOI are determined by the
desired load and θ50 at a given operating point. The objective is then to determine
the pilot-injection variables r p and d SOI . The influence of r p and d SOI on engine
performance is studied in the following sections.

143
Chapter 8. Pilot Injection

Table 8.1 The investigated operating points. The notion x : y : z indicates that
the corresponding parameter was swept from x to z in steps of y. Every combina-
tion of r p and d SOI was tested.

OP 1 OP 2

p IMEPg [bar] 5 10
Nspeed [rpm] 1200 1200
λ [-] 2.5 1.8
r EGR [-] 0.15 0.25
θ50 [CAD] 6 10
r p [-] 0 : 0.125 : 0.5 0 : 0.075 : 0.3
d SOI [CAD] 12.5 : 12.5 : 50 12.5 : 12.5 : 50
p rail [bar] 800 800

It would have been more convenient to use the injected fuel masses as input
variables as opposed to injection durations. Both for increased physical unders-
tanding and the more direct connection between the injected fuel amount and
the combustion processes. The reason for not doing so was due to the lack of
injector characteristics to compute the injection duration for a given demanded
fuel amount at the time of this study.

8.2 Experimental Characterization

Engine-output characteristics with respect to different pilot-injection config-
urations were investigated by varying r p and d SOI whilst θ50 and p IMEPg were kept
m tot
constant using θSOI and θDOI , at the operating points (OP) presented in Table 8.1.
The reason for keeping θ50 and p IMEPg constant was to exclude the pilot effect
on these variables, and to be able to answer the question of how the controller
should adjust a pilot injection for given θ50 and p IMEPg set points.
At each operating point, d p max , the gross indicated efficiency η GIE , NOx , un-
burned hydrocarbons (HC) and soot emission levels were measured in steady
m
state during 1000 cycles. The partial derivatives of θ50 with respect to θSOI and
p
θSOI were also investigated. The data obtained from these experiments and their
implications for controller design are presented and discussed in the following
sections.

144
 8.2 Experimental Characterization

d p max [bar/CAD], p IMEPg = 5 bar d p max [bar/CAD], p IMEPg = 10 bar

20 20
40 40
d SOI [CAD]

d SOI [CAD]
30 15 30 15

20 20
10
10
0 0.2 0.4 0 0.1 0.2
r p [-] r p [-]

Figure 8.4 The relation between the pilot-injection variables and d p max at
p IMEPg = 5, 10 bar. It is clear that r p can be used to control d p max since d p max
decreased with r p . The d p max controllability increased when d SOI was reduced.

Maximum Pressure-Rise Rate d p max

The measured influence of r p and d SOI on d p max is presented in Fig. 8.4. It can
be seen that d p max decreased with r p which makes r p a candidate for control-
ling d p max . At a certain r p , however, additional pilot fuel does not affect the
main-injection combustion rate, and the effect of r p saturates. The d p max con-
trollability varied with d SOI and was higher for small d SOI .
These results agree with previous research which states that the reactions of
the injected pilot fuel decrease the ignition delay of the main-injection, which in
turn decreases the heat-release rate. This effect is enhanced for larger pilot-fuel
amounts. This explanation is supported by the computed main-injection igni-
tion delay

τm = θ10 − θSOI
m
[ms] (8.7)

which is presented in Fig. 8.5, where it can be seen that τm and d p max correlate.
The pilot effect weakened with increased d SOI . This could be explained by a
m
more dilute pilot mixture at θSOI , with a lowered temperature increase as a result.
An increased d SOI would also result in more diverse fuel-spray targets for the two
injections. Early pilot injections put more fuel in the crevice volume outside of
the piston bowl where the main-injection is targeted. The resulting spatial sepa-
ration of the injections could explain the observed trends. This was suggested by
optical engine data obtained from a similar heavy-duty engine setup, see [Lönn
et al., 2017]. Furthermore, the results in Figs. 8.4 and 8.5 indicate a significant
trade-off between obtainable τ and d p max , which could be problematic if a long
τ is required.

145
Chapter 8. Pilot Injection

τm [ms], p IMEPg = 5 bar τm [ms], p IMEPg = 10 bar

2 2
40 40
d SOI [CAD]

d SOI [CAD]
30 1.5 30 1.5

20 20
1 1

0 0.2 0.4 0 0.1 0.2

r p [-] r p [-]

Figure 8.5 The relation between the pilot-injection variables and τm at p IMEPg =
5, 10 bar. The main-injection ignition delay correlated with d p max in Fig. 8.4.

Indicated Efficiency
The pilot-injection impact on efficiency was evaluated by computing the gross
indicated efficiency
p IMEPg Vd
η GIE = (8.8)
m f Q LHV
where m f is the injected fuel amount, computed from the average fuel-flow du-
ring the experiment. The results presented in Fig. 8.6 show that η GIE had a shal-
low maximum when d SOI was short. This trend was more significant for OP 1,
where η GIE was more sensitive to the pilot configuration. Another visible trend is
that η GIE decreased when d SOI increased, and that this effect was stronger when
r p was higher. These results show that it could be efficient to have a pilot injec-
p
tion but that this effect is reversed for very early θSOI .

HC
The measured unburned hydrocarbon (HC) emission levels are presented in Fig.
8.7. HC emissions increased steeply as d SOI and r p were simultaneously incre-
ased. The explanation for this could be that the pilot fuel was injected into the
crevice regions outside of the piston bowl and did not burn completely due to
wall-cooling effects and too lean conditions. This effect is clearer at p IMEP = 5
bar where in-cylinder temperatures were lower. The decrease in combustion eff-
iciency indicated by the HC-emission increase could explain the η GIE decrease
observed in Fig. 8.6.

NOx
The measured NOx emission levels are presented in Fig. 8.8, where it can be
seen that NOx emissions mainly depended on r p and decreased with an in-

146
 8.2 Experimental Characterization

ηGIE [-], p IMEPg = 5 bar ηGIE [-], p IMEPg = 10 bar

0.54 0.54

40 0.52 40 0.52
d SOI [CAD]

d SOI [CAD]
30 0.5 30 0.5

20 20
0.48 0.48

0 0.2 0.4 0 0.1 0.2

r p [-] r p [-]

Figure 8.6 The relation between the pilot-injection variables and η GIE at
p IMEPg = 5, 10 bar. The results show that η GIE had a shallow maximum when the
pilot was close to the main injection. Another visible trend is that η GIE decreased
when d SOI increased, and that this effect was stronger for larger r p .

creased r p . A hypothetical explanation for the decrease in NOx is the decrease

in τm with r p . The decrease in τm resulted in a decrease in heat-release rate
and in-cylinder pressure, which indicates lowered peak temperatures and de-
creased NOx -formation rates. Lower peak temperatures also results in reduced
heat-transfer losses, which could explain the η GIE increase with r p in Fig. 8.6.

Soot
Figure 8.9 shows that measured soot levels increased with r p and more so when
d SOI was small. An explanation for this is that the ignition delays for the fuel in-
jections were minimized at these operating points. A decreased air/fuel mixing
time results in richer combustion and increased soot formation. The increase in
soot emissions was not as large for longer d SOI , which means that d p max could
be reduced with a lower soot-emission penalty. Similar NOx and soot-emission
trends were presented in [Manente et al., 2009, 2010b].

Combustion-Timing Controllability
m p
The θ50 controllability was investigated by varying θSOI and θSOI in open loop at
m p
the operating points in Table 8.1. For each operating point, θSOI and θSOI were
varied individually in square-wave sequences with an amplitudes of 1 CAD and
m,p
a period of 25 cycles during 500 cycles. The partial derivatives ∂θ50 /∂θSOI were
then computed.
Computed partial derivatives in Figs. 8.10 and 8.11 indicate that θ50 is con-
m p
trolled by θSOI , and that the controllability decreased slightly with r p . The θSOI
p
effect was an order of magnitude smaller, where θSOI affected θ50 more when r p

147
Chapter 8. Pilot Injection

HC [ppm], p IMEPg = 5 bar HC [ppm], p IMEPg = 10 bar

800 800
40 40
d SOI [CAD]

d SOI [CAD]
600 600
30 30
400 400
20 20
200 200

0 0.2 0.4 0 0.2 0.4

r p [-] r p [-]

Figure 8.7 The relation between the pilot-injection variables and HC at

p IMEPg = 5, 10 bar. The HC emission levels increased steeply as d SOI and r p were
simultaneously increased. The explanation for this could be that the pilot fuel
was injected into the crevice regions outside of the piston bowl, resulting in in-
complete combustion. The decrease in combustion efficiency indicated by the
HC emission-level increase explains the η GIE decrease observed in Fig. 8.6.

was increased. A hypothetical explanation is that the main injection initiated the
high-temperature reactions with corresponding heat-release rate. Similar trends
were shown in [Hasegawa and Yanagihara, 2003; Manente et al., 2009].
Another observation is that the pilot injection linearizes the relation between
m
θSOI and θ50 , due to its reduction of τ. This argument is supported by the fin-
m
dings in Chapter 6, where τ contributed to a nonlinear relation between θSOI and
θ50 . A pilot injection therefore increases the controllable region in Fig. 6.2, which
facilitates θ50 control.

Summary
The experimental findings show that r p can be used to control d p max . The choice
of r p gives a trade off between obtainable τ and d p max and a trade off between
soot and NOx emissions. A controller-design approach would be to let a fast
cycle-to-cycle controller decide r p to obtain acceptable d p max , and then let the
gas-exchange system set suitable boundary conditions in terms of EGR ratio and
intake temperature that simultaneously allow for low d p max and high τ.
The efficiency η GIE was shown to vary slightly with r p where η GIE decreased or
increased depending on d SOI . This is believed to be linked to the observed trends
in NOx and HC-emission levels as discussed previously. This hypothesis should
be confirmed with optical experiments.
The injection separation d SOI was shown to be a trade-off between high
d p max controllability and η GIE with shorter d SOI , and simultaneously low NOx

148
 8.3 Controller Design

NOx [ppm], p IMEPg = 5 bar NOx [ppm], p IMEPg = 10 bar

1,400 1,400

40 1,200 40 1,200
d SOI [CAD]

d SOI [CAD]
1,000 1,000
30 30
800 800
20 20
600 600
0 0.2 0.4 0 0.1 0.2
r p [-] r p [-]

Figure 8.8 The relation between the pilot-injection variables and NOx at
p IMEPg = 5, 10 bar. It can be seen that NOx mainly depended on r p and decre-
ased with an increased r p . A hypothetical explanation for this is the decrease in
τm with r p . The resulting decrease in heat-release rate and in-cylinder pressure
with τm indicate lowered peak temperatures and decreased NOx -formation rates.

m
and soot emissions with increased d SOI . The experiments also showed that θSOI
should be used to control θ50 .

8.3 Controller Design

The controller-design objective was to keep θ50 at a predefined value, and main-
tain d p max below an upper bound. The experimental results above showed that
m
d p max can be controlled with r p while θ50 is mainly affected by θSOI . The injec-
tion separation was chosen to be constant and low, d SOI = 20 CAD, to maintain
high η GIE and high d p max controllability. In order to keep the ignition delay τm
as long as possible under these circumstances, it was decided to keep r p low. The
tot
engine load was controlled with θDOI in a separate closed loop using feedforward
and PI control with constant p rail .
The model-complexity needed to physically model the relation between pi-
lot injection and pressure-rise rate is fairly involved. An empirical modeling ap-
proach was therefore adopted, where experimental data provided an approx-
imate linear cycle-to-cycle model. The state x and input u are introduced ac-
cordingly
¡ ¢T
x(k) = θ50 (k) d p max (k) r p (k)
¡ m ¢T (8.9)
u(k) = ∆θSOI (k) ∆r p (k)
where k denotes the cycle index, and ∆ is the forward-difference operator. The
pilot ratio was introduced as a state to keep track of its absolute value.

149
Chapter 8. Pilot Injection

soot [mg/m3 ], p IMEPg = 5 bar soot [mg/m3 ], p IMEPg = 10 bar

150

40 4 40
d SOI [CAD]

d SOI [CAD]
100

30 30
2 50
20 20

0 0.2 0.4 0 0.1 0.2

r p [-] r p [-]

Figure 8.9 The relation between the pilot-injection variables and soot at
p IMEPg = 5, 10 bar. Note the difference in z-axis scale. Soot levels increased with r p
and more so when d SOI was small. An explanation for this is that the ignition de-
lays for the fuel injections were minimized at these operating points which decre-
ased air/fuel mixing time and resulted in richer combustion with increased soot
formation.

∂ θ 50 / ∂ θ m
SOI
[-], p IMEPg = 5 bar ∂ θ 50 / ∂ θ m
SOI
[-], p IMEPg = 10 bar
1.4 1.4

40 1.2 40
d SOI [CAD]

d SOI [CAD]

1.2

30 1 30
1
20 20
0.8
0.8
0 0.2 0.4 0 0.1 0.2 0.3
r p [-] r p [-]

Figure 8.10 The computed gain from θSOI m to θ at p

50 IMEPg = 5, 10 bar. The gain
m
from θSOI was high for all injection configuration and decreased as r p was incre-
ased.

The cycle-to-cycle model

x(k + 1) = Ax(k) + B u(k) + v (k)

(8.10)
y(k) = C x(k) + e(k)

was then formulated, where the assumption of a static relation between u and x

150
 8.3 Controller Design

p p
∂ θ 50 / ∂ θ SOI [-], p IMEPg = 5 bar ∂ θ 50 / ∂ θ SOI [-], p IMEPg = 10 bar
0.2 0.2

40 0.1 40 0.1
d SOI [CAD]

d SOI [CAD]
30 0 30 0

−0.1 −0.1
20 20

−0.2 −0.2
0 0.2 0.4 0 0.1 0.2 0.3
r p [-] r p [-]

p
Figure 8.11 The computed gain from θSOI to θ50 at p IMEPg = 5, 10 bar. The
gain was low for all injection configurations. When comparing these trends with
m should be used to control θ .
Fig. 8.10, it is clear that θSOI 50

gives A = I 3x3 , and B contains the partial derivatives of x with respect to u

 ∂θ ∂θ50 
50
 ∂θSOI ∂r p   
  1 −3
 
B = ∂d p max ∂d p max  = −0.5 −25 (8.11)
 
 ∂θ ∂r p  0 1
 SOI 
0 1

From the experimental results in figures 8.10 and 8.11, it was found that
∂θ50 /∂θSOI = 1. The partial derivative ∂θ50 /∂r p = −3 was extracted from
Fig. 8.5. A decrease in θSOI resulted in an increased τm which in turn increa-
sed d p max , which explains ∂d p max /∂θSOI = −0.5. The d p max data in Fig. 8.4 gave
∂d p max /∂r p = −25. The matrix C in (8.10) is given by I 3x3 , since the first two
states can be computed directly using in-cylinder pressure measurements and
heat-release analysis. In order to incorporate model uncertainty and stochas-
tic cycle-to-cycle variation, zero-mean Gaussian processes v (k) and e(k) were
introduced with covariance matrices Q v and Q e

¡ ¢
Q v = E v (k)v (k)T
¡ ¢ (8.12)
Q e = E e(k)e(k)T

151
Chapter 8. Pilot Injection

Table 8.2 The chosen MPC weights.

ω1 ω2 ω3 ω4 ω5 ρ
100 0.01 25 6000 8000 1

Model Predictive Control

A linear-system model allows for linear MPC, which is suitable design choice for
output-constraint handling. The MPC problem was formulated according to
Hp
X r
minimize ω1 ||θ50 (k) − θ50 (k)||22 + ω2 ||θSOI
m
(k)||22 + ω3 ||r p (k)||22
m
θSOI , rp k=1
HX
c −1
+ ω4 ||∆θSOI (k)||22 + ω5 ||∆r p (k)||22 + ρ²2 (8.13)
k=0
subject to x(k + 1) = Ax(k) + B u(k) for k = 0, . . . , Hc − 1
x(k + 1) = Ax(k) for k = Hc , . . . , H p − 1
x(0) = x 0
0 ≤ r p (k) ≤ r pc
c
d p max (k) ≤ d p max +²
²≥0
where r pc and d p max
c
are upper bounds for r p and d p max . These were set to 0.3 and
8 bar/CAD, respectively. The variable ² is a cost variable that penalizes violation
of the d p max inequality constraint, and was introduced to ensure existence of fe-
asible solutions. The positive weights ωi and ρ determine the controller priority
and were tuned to obtain adequate closed-loop response times subject to over-
shoot and enhanced cycle-to-cycle variation. The weights used are presented in
Table 8.2. The horizons were chosen according to H p = 16 and Hc = 8. Average
computation times for solving (8.13) were 3 ms for one cylinder.

Kalman Filter
State estimation was used to handle cycle-to-cycle variation in d p max and θ50 . In
order to estimate x from y, u and (8.10), a stationary Kalman filter was used. The
Kalman-filter state estimate x̂ was updated recursively according to
x̂(k + 1) = A x̂(k) + B u(k) + K (y(k) −C ŷ(k)) (8.14)
where the Kalman-filter gain K was given by the steady-state solution to Algori-
thm 1 in Chapter 3
K = APC T (Q e +C PC T )−1
(8.15)
P = AP A T +Q v − APC T (Q e +C PC T )−1C P A T

152
 8.4 Controller Evaluation

The covariance matrices used

µ ¶ µ ¶
25 0 10 0
Qe = , Qv = (8.16)
0 800 0 50

were chosen to get sufficient measurement-noise attenuation. The choice Q v2,2 >
Q v1,1 reflects an increased model uncertainty in d p max . The Kalman filter only
considered the first two state equations in (8.10).

8.4 Controller Evaluation

Controller performance was evaluated experimentally during set-point chan-
c
ges in θ50 and p IMEPg , as well as during changes in d p max and Nspeed . Indica-
ted mean-effective pressure and p rail were controlled using PI controllers and
feedforward, whilst Nspeed was changed by adjusting the engine-brake speed set
point.
Input and output data for one cylinder are presented during a sequence of
r
θ50 step changes in Fig. 8.12, where the θ50 response time was in the range of 3
cycles. In the upper diagrams, x̂ is indicated in black together with y, which is
indicated in gray. The ignition delay and d p max increased as θ50 was advanced.
c
This forced r p to increase for d p max = 8 bar/CAD to be fulfilled. This controller
r c
behavior allows for a more advanced θ50 for a given d p max . The pilot ratio was
r
then decreased when θ50 was delayed due to its absolute-value cost.
In Fig. 8.13, input and output data are presented during step changes in
c c
d p max . The response time of d p max during a negative d p max step change was ap-
c
proximately 2 cycles, while the response time of d p max during a positive d p max
step change was approximately 10 cycles. This was due to the high penalty of vio-
c m
lating d p max . The main-injection timing θSOI adjusted for variation in θ50 caused
m
by changes in τ .
r
The closed-loop response during p IMEPg changes are presented in Fig. 8.14,
tot m
where the p IMEPg response time was approximately 20 cycles. As θDOI varied, θSOI
was adjusted to keep θ50 constant. The pilot ratio made minor adjustments to
c
keep d p max below d p max . The small variation in r p indicates that d p max was
tot
not very sensitive to changes in θDOI at this operating point. Input-signal os-
cillation can be observed at cycles 200 and 500 in Fig. 8.14, indicating that the
control-action weights could be increased for more robust performance.
System response to Nspeed changes are presented in Fig. 8.15. Here, the con-
m c r
troller had to increase r p and advance θSOI in order to fulfill d p max and θ50 when
Nspeed was increased. It can also be seen that the d p max noise level decreased
with Nspeed , which indicates a reduced cycle-to-cycle variation with Nspeed .

153
Chapter 8. Pilot Injection

d p max [bar/CAD]
14 15
12

θ 50 [CAD]
10 10
8 5
6
4 0
400 600 800 400 600 800
cycle [-] cycle [-]

−6 0.4
θ SOI [CAD]

−8 0.3

r p [-]
−10 0.2
−12 0.1
−14 0
400 600 800 400 600 800
cycle [-] cycle [-]

1.04 7.4
p IMEPg [bar]

1.02 7.2
[ms]

1 7
DOI

0.98 6.8
θ tot

0.96 6.6
0.94 6.4
400 600 800 400 600 800
cycle [-] cycle [-]

Figure 8.12 Input and output data during a sequence of θ50 r step changes. In

the upper diagrams, x̂ is presented in black together with y which is presented in

c
gray. The pilot ratio was forced to increase as θ50 was advanced to fulfill d p max =8
bar/CAD. As θ50 was delayed, r p was decreased due to its absolute-value cost.

8.5 Discussion and Conclusions

The results indicate that the designed controller was successful in maintaining an
upper bound for d p max whilst θ50 was kept at a predefined value. Both in steady
r
state and during load and speed transients. The response time to changes in θ50
c
was 3 cycles, and 2 cycles for a negative change in d p max . There were cases of sig-
nificant cycle-to-cycle variation in u for some of the experiments, see Figs. 8.14
and 8.15. This variation could be decreased by increasing the control-action wei-
ghts in (8.13).
The pressure-rise rate d p max was in this chapter treated as a stochastic sig-
nal whose mean value was to be controlled below an upper limit. The validity of
c
this treatment could of course be questioned since d p max levels above d p max will
c
occur even if the mean d p max level is kept below d p max . A more sophisticated de-

154
 8.5 Discussion and Conclusions

d p max [bar/CAD]
12 20
15

θ 50 [CAD]
10
10
8
5
6 0
200 400 600 200 400 600
cycle [-] cycle [-]

−9 0.4
θ SOI [CAD]

0.3
−10

r p [-]
0.2

−11 0.1
0
200 400 600 200 400 600
cycle [-] cycle [-]

1.15 7.5
p IMEPg [bar]
[ms]

1.1 7
DOI

1.05 6.5
θ tot

1 6
200 400 600 200 400 600
cycle [-] cycle [-]

c
Figure 8.13 Input and output data during step changes in d p max . The response
c
time of d p max during a negative d p max step change was 2 cycles, while the re-
c
sponse time of d p max during a positive d p max step change was 10 cycles. The
m
main-injection timing θSOI adjusted for variations in θ50 caused by changes in
rp .

sign approach would be to incorporate the statistical distribution of d p max in the

c
controller design, and in this way control the frequency or probability of d p max
violation as presented in [Jones and Frey, 2015].
It is possible that comparable performance could be obtained with a simpler
controller structure. For instance by handling the d p max limit as a set-point pro-
blem as presented in [Ott et al., 2013]. This would however demand additional
heuristic logic, and the framework would not be as general and easily expanda-
ble if more states and inputs were to be added to the control problem.
Suggested future work consists of investigating higher engine loads and the
controller compatibility with a gas-system controller that adjusts the engine
c
intake conditions to maximize τ for a given d p max . It would also be interes-
ting to generalize the control problem to incorporate more than two injections.

155
Chapter 8. Pilot Injection

d p max [bar/CAD]
12 14
10 12

θ 50 [CAD]
8 10
6 8
4 6
2 4
200 400 600 800 200 400 600 800
cycle [-] cycle [-]

−10 0.4
θ SOI [CAD]

−11 0.3

r p [-]
−12 0.2
−13 0.1
−14 0
200 400 600 800 200 400 600 800
cycle [-] cycle [-]

p IMEPg [bar]
1.2 8
DOI
[ms]

1 6
θ tot

0.8 4
200 400 600 800 200 400 600 800
cycle [-] cycle [-]

Figure 8.14 Input and output data during p IMEPg set-point changes. The re-
sponse time for p IMEPg was approximately 20 cycles. The pilot ratio was only
doing minor adjustments in order to keep d p max below the upper limit. This indi-
tot at this operating point.
cates that d p max was not very sensitive to changes in θDOI

Triple-injection strategies has previously been suggested for PPC operation [Ma-
nente et al., 2010a].
The injection separation d SOI was shown to impact both emissions and effi-
ciency, but was in this work held constant for simplicity. A suggested extension
of the controller presented here is therefore to incorporate d SOI as an input va-
riable in the controller design. This was for instance done in [Yang et al., 2017] to
control soot-emission levels.

156
 8.5 Discussion and Conclusions

d p max [bar/CAD]
20
10
15

θ 50 [CAD]
8 10
5
6
0
600 700 800 900 600 700 800 900
cycle [-] cycle [-]

−8 0.4
θ SOI [CAD]

−10 0.3

r p [-]
−12 0.2
−14 0.1
−16 0
600 700 800 900 600 700 800 900
cycle [-] cycle [-]

1,600
Nspeed [rpm]

1,500
1,400
1,300
1,200
600 700 800 900
cycle [-]

Figure 8.15 Input and output data during Nspeed set-point changes. The con-
m in order to fulfill d p c
troller increased r p and decreased θSOI r
max and track θ50
when Nspeed was increased. It can also be seen that the d p max -noise level de-
creased with Nspeed .

157
9
Low-Load Control

9.1 Introduction
To obtain reliable low-load operation is one of the major challenges with ga-
soline PPC, both due to high cycle-to-cycle variation and low combustion ef-
ficiency. Advantages with a high octane number (ON) have mainly been obse-
rved at mid-to-high engine load. With a high ON, the in-cylinder conditions at
low-load operation result in too long ignition delays for combustion to complete
during the closed part of the engine cycle [Fieweger et al., 1997]. The issue of
having too long ignition delays was discussed in Chapter 6, where elevated in-
take temperatures were needed to achieve θ50 close to TDC, see Fig. 6.2.
Incomplete combustion and misfire result in reduced engine efficiency and
increased HC and CO emission levels. Homogeneous-reactor simulations have
shown that high levels of HC and CO are a result of lean mixtures at too low tem-
perature [Kim et al., 2008]. Fuel reactivity is also an important factor. Manente et
al. [2010a] showed that as the fuel ON increased from 70 to 100, the PPC low-load
limit, defined as the p IMEP at the fuel-ignitability limit, increased from 3 to 15 bar
at atmospheric intake conditions.
Results in [Kalghatgi et al., 2006] and [Weall and Collings, 2009] showed that
a remedy to the PPC low-load issue is to increase in-cylinder temperature and
equivalence-ratio stratification levels, since combustion initiated in rich regions
with high temperature aids combustion of the overall cylinder charge. Previous
studies presented in [Borgqvist et al., 2012; 2013] aimed to improve the gaso-
line PPC low-load efficiency using variable-valve actuation. These results showed
that negative valve overlap and re-breathing in combination with a split-main in-
jection strategy were able to increase the low-load performance by increasing the
temperature of trapped residual gases. In [Solaka et al., 2012], the low-load limit
of a single-cylinder light-duty engine was extended down to p IMEPg = 2 bar, using
boosted intake air. The absolute intake pressure needed at this load was appro-
ximately 2 bar for the fuels with the highest ON (88.6 and 87.1). Other actuator
options for improved low-load operation include variable compression ratio [Ha-
raldsson et al., 2002] and fast thermal management [Martinez-Frias et al., 2000].

158
 9.2 Low-Load Experiments

This chapter investigates how fuel injection and intake conditions should be
chosen for improved PPC efficiency, with the objective of extending the operation
range towards lower loads. Section 9.2 presents experimental data for p IMEPn ∈
[1, 5] bar. This data were then used to suggest a simple controller that acts to
minimize the ignition delay τ during low-load operation, see Sec. 9.3.
Section 9.4 shows how the controller-design choices affect fuel consumption
during transient operation.

9.2 Low-Load Experiments

This section presents experimental data for p IMEPn ∈ [1, 5] bar. The data describe
combustion sensitivity to injection timing, pilot injection and intake conditions.

Injection Timing
The combustion timing θ50 has to be phased shortly after TDC for work ouput
to be maximized and temperatures to be sufficient for the chemical reactions to
complete. Experimental θ50 data are presented in Fig. 9.1 as a function of θSOI for
three different fuel-injection durations θDOI and p rail = 800 bar.
The gain from θSOI to θ50 was positive for θSOI > −20 CAD. As θSOI decrea-
sed from this point, the increase in τ with θSOI became more significant, and the
gain from θSOI to θ50 decreased. The gain even became slightly negative for very
early θSOI , as the increase in τ exceeded the decrease in θSOI . Similar θ50 / θSOI
characteristics were observed in Chapter 6, Fig. 6.2, with the difference that the τ
model in Chapter 6 was unable to capture the negative-gain region for early θSOI .
A physical explanation for a negative gain was given in [Kalghatgi et al., 2006],
where similar experimental results were presented: An increase in τ gives more
time for fuel-air mixing, which in turn gives leaner mixtures and even longer τ.
The model in Chapter 6 did not include the time-resolved equivalence ratio his-
tory of the fuel/air mixture when computing the accumulated reactivity, and was
therefore not able to capture the negative-gain effect.
The data in Fig. 9.1 provide a lower bound for which θSOI can be used to ef-
fectively control θ50 with a linear controller. With integral action and a too low
infeasible set point, the controller would keep advancing θSOI and instead in-
crease set-point deviation. It was therefore decided to saturate θSOI at -20 CAD
to maintain a positive gain. In this way, the controller could obtain the wanted
set-point or the earliest obtainable θ50 , without risking unnecessarily long τ.

Pilot Injection
It was previously shown in Chapter 8 that a pilot injection reduces τ of the main
injection. A decrease in main-injection τ makes it possible to obtain a more ad-
vanced θ50 . This pilot-injection effect has previously been observed for conven-
tional diesel combustion [Osuka et al., 1994; Macmillan et al., 2009].

159
Chapter 9. Low-Load Control

20
θ DOI = 0.65 [ms]
θ DOI = 0.70 [ms]
θ DOI = 0.75 [ms]
15
θ 50 [CAD]

10

−30 −25 −20 −15 −10

θ SOI [CAD]

Figure 9.1 Combustion timing θ50 as a function of fuel-injection timing θSOI for
three different fuel-injection durations θDOI . For injections closer to TDC, the gain
between θSOI and θ50 was positive. As θSOI was advanced, the gain decreased and
became slightly negative. The negative gain was more significant for the shorter
injection durations.

p
Experiments with different pilot- and main-fuel injection durations, θDOI and
m p
θDOI , were conducted to investigate the θDOI effect on the gross indicated effi-
ciency

p IMEPg Vd
η GIE = (9.1)
m f Q LHV

The combustion timing θ50 was kept as close to TDC as possible, whilst the pi-
lot injection was positioned 10 CAD prior to the main injection. A small separa-
tion between pilot and main was previously shown to minimize τ, see Fig. 8.5.
m
Level curves of p IMEPn (blue, dashed) and η GIE (red, solid) as a function of θDOI
p
and θDOI are presented in Fig. 9.2. Figure 9.2 shows that for a given p IMEPn , η GIE
p
could be increased by having a longer θDOI . This effect became less significant
for higher p IMEPn . The pilot-injection effect on heat-release rate is presented in
Fig. 9.3, where it can be seen that the pilot injection both reduced τ and increased
the heat-release rate. From these results it was concluded that a pilot injection
should be used to aid main-injection ignition and allow for a more advanced θ50 .

160
 9.2 Low-Load Experiments

0.4
0. 6
0.38

2. 5
2

0.
5

3
0.35 0.5
5

0.33 0.45
θ DOI [ms]

0.4
1

1.

2. 5
5
0.3

2
0.35
p

0.28 0. 4
5

0.3
0.
0.25
4
0.5

0.2
1.5
0.23 0.5 p IMEPn

2.5
1

2
ηGIE
0.2
0.6 0.63 0.65 0.68 0.7 0.73 0.75 0.78 0.8
θm
DOI
[ms]

Figure 9.2 Level curves of p IMEPn (blue, dashed) and η GIE (red, solid) as a func-
m and θ p . For a low p
tion of θDOI DOI IMEPn , η GIE could be increased by having a
p
longer θDOI . This effect became less significant for higher p IMEPn .

80

60
p [bar]

40

20

0
−30 −20 −10 0 10 20 30
θ [CAD]

Figure 9.3 In-cylinder data comparing heat-release and pressure with and wi-
thout a pilot injection. With the same main-injection timing, τ was reduced by the
pilot injection, which gave an advanced θ50 and an increased heat-release rate.

161
Chapter 9. Low-Load Control

ηGIE [-]

0.3

0.4
0.25

0.2 0.35
φ [-]

0.15
0.3

0.1
0.25

710 720 730 740 750 760 770 780 790

TθSOI [K]

Figure 9.4 Experimental η GIE as a function of equivalence ratio φ and tempera-

ture at θSOI . For low φ, η GIE increased with temperature. These data were used to
model η NIE in the MPC simulations.

Intake Conditions
The gas-system valve positions θHP and θcool were varied at the following injec-
tion durations
m
θDOI = {0.65, 0.7, 0.75, 0.85, 1.05} [ms] (9.2)
to investigate how the intake conditions affect η GIE . The combustion timing was
kept in the interval 0-5 CAD and no pilot injection was used. Computed η GIE du-
ring these experiments are presented as a function of global equivalence ratio φ
and temperature at θSOI , TθSOI in Fig. 9.4. The temperature TθSOI was computed
by assuming adiabatic compression
µ ¶
V (θIVC ) γ−1
TθSOI = Tin (9.3)
V (θSOI )
In Fig. 9.4, the efficiency η GIE decreased steeply when φ and TθSOI were simulta-
neously reduced. It can also be seen that η GIE was more sensitive to TθSOI at low
φ, and could be increased significantly by increasing TθSOI . The effect of increa-
sing Tin can also be seen in Fig. 9.5, where the increase in Tin decreased τ and
increased the heat-release rate.

162
 9.2 Low-Load Experiments

80
Tin = 317 K
Tin = 293 K
p [bar] 60

40

20

0
−30 −20 −10 0 10 20 30
θ [CAD]

Figure 9.5 The effect of elevating the intake temperature. This was done by
opening the high-pressure EGR valve and closing the cool thermal-management
valve. An increased temperature gave a decrease in τ as well as an increased
m
heat-release rate, which indicates an improved combustion efficiency since θDOI
was kept constant.

Varying θHP and θcool also affected pumping losses p PMEP , which is the in-
dicated mean-effective pressure during the gas-exchange strokes. The relation
between p PMEP and the intake and exhaust manifold pressures, p in , p ex , is pre-
sented in Fig. 9.6, where the symbols ° and 4 relate the data in Figs. 9.4 and 9.6.
Pumping losses correlated with p in − p ex as expected. Moreover, the symbols in-
dicate that p PMEP in Fig. 9.6 correlated with TSOI in Fig. 9.4. This was because ope-
ning θHP not only heated the intake charge, it also elevated the intake-manifold
pressure, and in that way reduced pumping losses. The data presented in Figs. 9.4
and 9.6 were used for efficiency optimization through simulation in the following
section.

163
Chapter 9. Low-Load Control

p PMEP [bar]
−0.24
1.28
−0.26
1.26
−0.28

1.24 −0.3
p ex [bar]

1.22 −0.32

−0.34
1.2
−0.36
1.18
−0.38
1.16
0.96 0.98 1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16 1.18
p in [bar]

Figure 9.6 Experimental p PMEP as a function of intake and exhaust-manifold

pressures, p in , p ex . These data were used to model η NIE in model predictive con-
trol simulations.

Optimal Gas-Exchange Actuation

The efficiency data in Figs. 9.4 and 9.6 were used together with the calibrated
gas-exchange-system model, presented in Sec. 2.3, to find efficiency-optimal
θcool and θHP actuation during p IMEPn set-point changes.
A model predictive control problem was formulated for this purpose
Hp
X
minimize ω1 |m f (k)| + ω2 ||1 − θcool (k)||22 + ω3 ||θHP (k)||22 (9.4)
θcool , θHP k=1


θHP
 θcool 
subject to θ l ≤  
 ∆θHP  ≤ θ
u

∆θcool
ẋ = f (x, θcool , θHP ), (2.22) − (2.37)
η GIE = g 1 (φ, TθSOI )
p PMEP = g 2 (p 1 , p 2 )

The objective of this controller is to minimize fuel consumption. This is repre-

164
 9.2 Low-Load Experiments

sented by the first cost term in (9.4). The other two terms penalize deviation from
suitable valve positions at mid-to-high engine load where intake-manifold tem-
perature should be kept low to reduce heat-transfer losses and the low-pressure
EGR path is preferred over the high-pressure EGR path. Valve positions θcool and
θHP are here normalized from 0 to 1, where 1 denotes fully open. The cost func-
tion is defined over a prediction horizon of H p engine cycles, where k denotes
cycle index. The cost function was minimized subject to the dynamics of the
gas-exchange system model in Sec. 2.3. Here, the state x include the pressures
and temperatures of the volumes in Fig. 2.3. The functions g 1 and g 2 were obtai-
ned by interpolating the data in Figs. 9.4 and 9.6.
The gas-exchange model in Sec. 2.3 was then used to simulate the MPC in
(9.4). During this experiment, p IMEPn was set to follow a set-point trajectory using
a PI controller. Combustion timing θ50 was also regulated through θSOI adjust-
ments. No combustion model was used since p IMEPn was given directly by g 1 and
g 2 . An ignition-delay model was however used to compute TθSOI . The nonlinear
MPC problem (9.4) was solved every simulated engine cycle using the MATLAB
nonlinear-optimization toolbox together with the ode23s solver to compute the
gas-system model output. Global solutions of (9.4) can not be guaranteed due to
the nonlinearity of the optimization problem. Solutions obtained were however
justified by confirming that they were consistent for different initial guesses.
Simulation results for three MPCs with different fuel-consumption weights
ω1 = 0, 1, 5, ω2 = ω3 = 10 and H p = 10 are presented in Fig. 9.7. As the cost for fuel
consumption was increased, the controller increased gas-exchange actuation by
closing θcool and opening θHP when p IMEPn was reduced. The controller was in
this way able to avoid the low-efficiency region in the φ − T diagram (see the
lower right subdiagram in Fig. 9.7). This lowered the injected fuel amount m f
needed at low load. Overshoot in p IMEPn was also slightly reduced.
Fuel efficient actuation of θcool and θHP in Fig. 9.7 was approximated with a
static φ-feedback law K (φ) µ ¶ µ ¶
θcool K 1 (φ)
= (9.5)
θHP K 2 (φ)
Figure 9.8 shows K coincidence with the actuated valve positions in Fig. 9.7.
There is some deviation between data and K 1 for higher φ. This deviation was
deliberately chosen to utilize the full range of θcool for reduced intake-manifold
temperatures at higher engine loads. The reason for approximating the MPC be-
havior was due to the inhibiting computation times needed to solve (9.4) online.

165
Chapter 9. Low-Load Control

80 6

p IMEPn [bar]
60
4
m f [mg] 40
2
20

0 0
0 2 4 6 8 0 2 4 6 8
time [s] time [s]

0.3
1
0.2

θ cool [-]
θ HP [-]

0.5
0.1

0 0
0 2 4 6 8 0 2 4 6 8
time [s] time [s]

800 0.5
0.4
TθSOI [K]

750 0.3
φ [-]

0.2

700 0.1
0
0 2 4 6 8 680 700 720 740 760 780
time [s] TSOI [K]

ω1 = 0 ω1 = 1 ω1 = 5

Figure 9.7 Simulated MPC output for different fuel-consumption weights ω1 .

With a larger weight, the controller avoids the low efficiency region in the φ − T
diagram. This behavior is indicated by the lower right subdiagram where the level
curves represent η GIE obtained from the data in Fig. 9.4.

9.3 Suggested Controller

The results presented suggest the following controller design:

• The combustion timing θ50 can be controlled by adjusting θSOI with a PI

r
controller. The set point θ50 should then be located close TDC in order to
avoid misfire. Furthermore, θSOI should be limited to the positive-gain re-
gion in Fig. 9.1, to ensure θ50 controllability and closed-loop stability.

166
 9.3 Suggested Controller

90
K 1 (φ)
K 2 (φ)
75 ∗
θcool
∗
θHP
60
θ HP / θ cool [deg]

45

30

15

0
0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24 0.26 0.28 0.3 0.32
φ [-]

Figure 9.8 The static feedback law K was used to approximate efficient MPC va-
lve positions in Fig. 9.7. It was decided to use the full range of θcool which gave a
∗
deviation between θcool and K 1 for higher φ.

• A pilot injection should be used to reduce τ for advanced θ50 and increased
heat-release rates. It was here decided to use a small pilot, located 10 CAD
prior to the main-injection. The reason the short separation time was to
minimize τ, see Fig. 8.5.

• Gas-exchange valve positions should be set according to the feedback law

K (φ) in Fig. 9.8 to increase intake temperature at low load. In this work, φ
was computed from intake-manifold conditions (2.27) and fuel-mass flows
from a calibrated injection-duration map.

The suggested feedback loop is presented in Fig. 9.9. The controller is fairly sim-
ple, and the decentralized controller design is easy to implement. The controller
would nevertheless be able to maximize efficiency according to the presented ex-
perimental results. The design of a centralized controller structure, that for exam-
ple utilizes ignition-delay and gas-exchange models, as presented in Chapter 7,
is suggested future work. Another possible extension to the controller in Fig. 9.9
is a method that experimentally identifies the θ50 gain, in order to adapt the θSOI
saturation limit.

167
Chapter 9. Low-Load Control

−1

θSOI
r
θ50 P
PI
θ50
Engine
θHP φ
θcool

K (φ)

Figure 9.9 The experimental results suggested the following low-load con-
trol strategy: A range-limited θ50 PI controller in combination with a static
gas-exchange controller with global φ as feedback variable. Load and rail pressure
are controlled separately using PI controllers with feedforward.

9.4 Controller Evaluation

The suggested controller was evaluated experimentally during a test cycle where
p IMEPn was set to follow set-point changes from 1 to 5 bar at Nspeed = 1200 rpm.
Four different controller-design cases were evaluated in order to compare con-
troller performance:

1. Cylinder-individual θ50 PI controllers.

2. Cylinder-individual θ50 PI controllers with θSOI saturation (> −20 CAD).

3. Case (2) with a pilot injection.

4. Case (3) with the gas-system feedback law K (φ).

For all cases, p IMEPn was controlled by keeping the rail pressure constant and
varying the main-injection duration with PI controllers and feedforward.
Experimental test-cycle results from one cylinder are presented in Fig. 9.10
for cases (1-4). Differences between the cases are more noticeable at low engine
load. The injection timing was not limited in case (1). This led to very early in-
r
jection timings when the earliest obtainable θ50 was larger than θ50 . The early
injection timing in case (1) resulted in a τ increase and a delayed θ50 as a result.
r
When limiting θSOI , as in case (2), θ50 was advanced and the deviation from θ50
was reduced.

168
 9.5 Conclusions

When a pilot injection was introduced in case (3), τ decreased, see the θSOI
and θ50 subdiagrams in Fig. 9.10. The reduction in τ allowed the controller to
r
keep θ50 closer to θ50 . In case (4), τ was further decreased as Tin and TθSOI in-
creased due to valve-position actuation. This resulted in smaller θ50 error at low
load.
In-cylinder data showing these trends more clearly are presented in Fig. 9.11
where in-cylinder pressure, injection current and heat-release rate are presented
at cycle 250. It can be seen that the gradual controller adjustments led to decrea-
sed τ. The heat-release rate also differed for the different cases where case 1, with
the longest ignition delay had the lowest heat-release rate.
Accumulated fuel consumption, computed from injection durations is pre-
sented in Fig. 9.12. The fuel-consumption rates differed more clearly at the
low-load operating points. Figure 9.12 shows that fuel consumption decreased
from case (1-4), with reductions from 2 to 9 %. The greatest reduction resulted
from the introduction of a pilot injection.

9.5 Conclusions
This chapter presented a control strategy for improved PPC performance at low
load. The combustion timing θ50 was only controllable with respect to θSOI in a
specific interval, see Fig. 9.1. It was therefore important keep θSOI in this interval
to maintain closed-loop stability with a θ50 -feedback controller.
A pilot injection was shown to increase the combustion efficiency at low load,
see Figs. 9.2 and 9.12. This was due to the decrease in τ, which led to an increased
heat-release rate and the possibility to further advance θ50 .
The problem of maximizing η GIE was formulated as an optimization problem
in the φ-T diagram, where η GIE could be increased by heating the inducted air
charge, see Fig. 9.4. Control of the intake conditions was achieved by varying θHP
and θcool according to a feedback law K (φ), obtained from MPC simulations.
An experimental controller evaluation showed that these findings could re-
duce fuel consumption from 2 to 9 %, see Fig. 9.12.

Summary of PPC Control

This chapter concludes the results related to PPC control. The following three
chapters of this thesis cover controller designs applicable to conventional
compression-ignition operation. A summary of the PPC results presented in
chapters 6 to 9 will be given in Chapter 13.

169
Chapter 9. Low-Load Control

1.2

p IMEPn [bar]
5

θ DOI [ms]
4 1
3 0.8
2 0.6
1 0.4
0
500 1,000 1,500 500 1,000 1,500
cycle [-] cycle [-]

20
−10

θ SOI [CAD]
15
θ 50 [CAD]

10 −20
5
−30
0
500 1,000 1,500 500 1,000 1,500
cycle [-] cycle [-]

25
20 310
θ HP [deg]

Tin [K]

15
300
10
5 290
0
500 1,000 1,500 500 1,000 1,500
cycle [-] cycle [-]

100 0.75
θ cool [deg]

75
0.5
φ [-]

50
0.25
25
0 0
500 1,000 1,500 500 1,000 1,500
cycle [-] cycle [-]

Case 1 Case 2 Case 3 Case 4

Figure 9.10 Experimental results for the four different controller cases. In case 1,
θSOI was not limited to the positive-gain region. This led to very early θSOI , and as
a result, a delayed θ50 (compare case 1 and case 2). With a pilot injection in case
3, τ decreased, see the θSOI and θ50 subdiagrams. Finally, in case 4, τ decreased
further as θHP was opened and θcool was closed.

170
 9.5 Conclusions

p IMEPn = 2 bar
80
Case 1
Case 2
Case 3
60 Case 4
p [bar]

40

20

0
−30 −25 −20 −15 −10 −5 0 5 10 15 20
θ [CAD]

Figure 9.11 In-cylinder data from cycle 950 in Fig. 9.10. The gradual controller
adjustments from case 1 to case 4 resulted in a decreased τ. It can also be seen
that the advanced θSOI in case (1) gave a reduced heat-release rate.

171
Chapter 9. Low-Load Control

·105
3.5
Case 1
3 Case 2 (−2%)
Case 3 (−8%)
Case 4 (−9%)
Fuel Consumption [mg]

2.5

1.5

0.5

0
0 200 400 600 800 1,000 1,200 1,400 1,600 1,800
cycle [-]

Figure 9.12 Accumulated fuel consumption for the four controller cases in
Fig. 9.10 where the fuel-consumption rate differed more at low load. Fuel con-
sumption decreased from case (1-4), with reductions from 2 to 9 % where the gre-
atest reduction came with the introduction of a pilot injection. Here, the injected
fuel mass was computed from fuel-injection durations. The reason for not using
the fuel-mass-flow meter was that the meter was mounted far from the engine,
and was therefore not reliable in transient operation.

172
10
Constraint Handling with
Multiple Injections

10.1 Introduction
Despite its lack of spatial information, the heat-release rate as a function of
crank-angle degree is an important variable when maximizing the thermody-
namic engine efficiency. When assuming a zero-dimensional, adiabatic model
with constant γ, the thermodynamic efficiency is maximized when heat is re-
leased instantaneously at TDC. This heat-release rate minimizes exhaust losses
without generating counterproductive pressure during the compression stroke,
and is equivalent to the ideal Otto cycle with constant-volume combustion at
TDC.
The optimal combustion timing is delayed to after TDC when heat-transfer
to cylinder walls is also accounted for. A delayed combustion timing reduces
in-cylinder temperature and heat-transfer losses. It also reduces peak in-cylinder
pressure which in turn reduces engine friction and heat-losses to crevice volu-
mes. A drawback with delaying the combustion timing is the resulting increase
in exhaust-gas energy. The optimal combustion timing is therefore a compromise
between these losses.
The optimal heat-release rate becomes more involved when constraint ful-
fillment is required. With constraints on maximum cylinder pressure and NOx
formation, motivated by mechanical tolerances and emission regulations, com-
bustion has to be delayed during the expansion stroke to reduce the in-cylinder
pressure and temperature. Furthermore, efficient aftertreatment-system perfor-
mance requires sufficient exhaust-gas temperature, which in turn demands in-
creased exhaust losses [Gieshoff et al., 2000; Katare et al., 2009]. As suggested by
these examples, there are also compromises between constraint fulfillment and
the thermodynamic efficiency.
Optimal heat-release rates with respect to constraints on in-cylinder pres-
sure, NOx formation and knock intensity were computed in [Eriksson and Sivert-

173
Chapter 10. Constraint Handling with Multiple Injections

sson, 2016; Guardiola et al., 2017]. In these studies, optimal heat-release rates
were found to be multimodal to reduce in-cylinder pressure and temperature.
With a single fuel injection and a production heavy-duty fuel-injection system,
however, the heat-release rate controllability is limited. The combustion timing
can be controlled more or less freely by varying the injection timing, but the
heat-release shape depends on rates of fuel injection, mixing and chemical reac-
tions. These rates can only be controlled partially by adjusting cylinder-mixture
properties and the fuel-injection pressure.
The degrees of freedom increase when multiple injections are used. It
was found in [Han et al., 1996] that two injections can be used to provide a
more distributed heat-release rate with reduced peak temperature and NOx
formation. This allows for a more advanced effective combustion timing
without NOx -constraint violation. Okamoto and Uchida [2016] presented a
multiple-injector strategy for heat-release shaping. Experimental results showed
that heat-release shaping could provide a 75 % reduction of NOx emissions
due to suppressed peak average temperature and pressure with maintained in-
dicated efficiency. An indicated specific-fuel consumption reduction of 12 %
was reported in [Dober et al., 2008] with maintained NOx emissions, using a
novel fuel-injection system and multiple injections. The use of post injections
for control of exhaust-gas temperature was demonstrated in [Zheng et al., 2005;
Castellano et al., 2013] among others.
Some insight to the potential efficiency benefit with multiple injections can
be gained by studying the simulated pressure curves in Fig. 10.1. Here, the dif-
ferent heat-release rates and injection configurations generate the same work
output, p IMEPn = 15 bar. Without a constraint on maximum pressure p max , the
optimal combustion timing is close to TDC with the gross indicated efficiency
η GIE = 0.5, see the upper subdiagram. When a p max constraint is imposed (lower
subdiagram), the combustion timing has to be delayed for constraint fulfillment.
This results in a 4.8 % η GIE decrease with one injection (blue). With two injec-
tions (red), however, the combustion timing is allowed to be advanced, and the
decrease in η GIE is only 0.8 %.
An extensive calibration effort is demanded in order to find efficiency-optimal
fuel-injection configurations that fulfill pressure, NOx and exhaust-temperature
constraints over the engine operating range. Optimal configurations are also sen-
sitive to hardware aging, fuel properties and variation in operating conditions.
Moreover, the engine should operate as close to the constraint as possible for
maximized engine efficiency.
This chapter therefore investigates the use of feedback control for automa-
tic fuel-injection adjustment and increased engine efficiency subject to specified
constraints. The suggested controller is a hybrid, multiple-input multiple-output
PI controller that utilizes feedback from in-cylinder pressure-sensor measure-
ments, an NOx -emission model functioning as a virtual sensor, and measured
exhaust temperature. The controller varies the number of injections and adjusts

174
 10.1 Introduction

Unconstrained p max
200

150 ηGIE = 0.5

p [bar]

100

50

0
−20 −10 0 10 20 30 40 50
θ [CAD]

Constrained p max , c p max = 125 bar

200
One Injection
η1GIE = 0.476 η2GIE = 0.496 Two Injections
150
p [bar]

100

50

0
−20 −10 0 10 20 30 40 50
θ [CAD]

Figure 10.1 Simulated pressure curves with and without constraints on p max .
Without a constraint on p max , the optimal combustion timing is close to TDC with
the gross indicated efficiency η GIE = 0.5, see the upper subdiagram. When a p max
constraint at 125 bar is imposed, the combustion timing has to be delayed (see
lower subdiagram). This results in a 4.8 % η GIE decrease, with one injection (1).
With two injections (2), the combustion timing can be advanced, and the decrease
in η GIE is only 0.8 %.

injection timings and durations depending on operating conditions. Moreover,

the presented controller uses the heat-release detection method presented in
Chapter 4 to distinguish heat-release rates from different injections. The desig-
ned controller behavior was motivated by zero-dimensional simulation experi-
ments that are also presented in this chapter. These simulation results show how
optimal combustion timings vary as a function of constraint limits.
The chapter is outlined as follows: The problem formulation is given in
Sec. 10.2. The zero-dimensional model used in simulation is presented together

175
Chapter 10. Constraint Handling with Multiple Injections

with simulation results in Sec. 10.3. The proposed controller design is then intro-
duced in Sec. 10.4. Experimental evaluation results are presented in Sec. 10.5,
where both transient and steady-state operation are evaluated. Controller per-
formance is also compared to that of a simpler single-injection controller. Finally,
discussion and conclusions are given in Secs. 10.6 and 10.7.

10.2 Problem Description

The objective of this chapter is to maximize the gross indicated efficiency

p IMEPg Vd
η GIE = (10.1)
m f Q LHV

The efficiency η GIE should be maximized whilst p max , formed NOx emissions and
the exhaust temperature Tex fulfill upper and lower bounds
c
p max ≤ p max
NOx ≤ NOcx (10.2)
c
Tex ≥ Tex

Upper bounds on p, d p/d θ and NOx are motivated by mechanical engine tole-
rances and legislated emission limits. The lower limit for Tex is introduced to gua-
r
rantee after-treatment system performance. The demanded work output p IMEPn
should also be delivered
r
p IMEPn = p IMEPn (10.3)
The case with up to two fuel injections is first considered. The optimization va-
1 2
riables are the injection timings, denoted θSOI , θSOI , and injected fuel masses, m 1f
and m 2f . The optimization problem can be simplified by assuming that the total
fuel mass m tot
f
r
is determined by p IMEPn . The ratio

m 1f
r= (10.4)
m tot
f

can then be optimized instead of m 1f and m 2f . This ratio was here limited to

0.5 ≤ r ≤ 1 (10.5)
1 2
to exclude redundant configurations. Injection durations θDOI and θDOI are de-
tot
termined by r , m f and p rail , with the use of an injector map M inj

¡ 1 ¢T
θDOI 2
θDOI = M inj (r, m tot
f , p rail ) (10.6)

176
 10.3 Simulation

The injection pulses are not allowed to overlap in order to ensure that the inj-
ector needle closes in-between injections. This imposes constraints on possible
fuel-injection timings and durations.
The optimization variable
¡ m p ¢T
u = θSOI θSOI r Q tot (10.7)

is related to the system output variables

¡ ¢T
y = η GIE p IMEPn p max NOx Tex (10.8)

through a non-trivial relation

y = f (u) (10.9)

We can now formulate the optimization problem as

maximize η GIE (u) (10.10)

u
c
subject to p max (u) ≤ p max
NOx (u) ≤ NOcx
c
Tex (u) ≥ Tex
r
p IMEPn (u) = p IMEPn
y = f (u)
u ∈U

where the set U denotes feasible fuel-injection configurations.

The most difficult part of solving (10.10) is to evaluate f for a given u.
In reality, f is determined by injector dynamics, turbulent combustion and
thermodynamic processes. The approach taken here was to model f with the
zero-dimensional model presented in Chapter 2, and then evaluate f over a grid
of inputs u. This was done both to find optimal system inputs u ∗ and to find
trends for how u ∗ depend on different constraints x c . Simulation results showing
u ∗ for different x c are presented in the following section. These results provide a
foundation for controller design in subsequent sections.

10.3 Simulation
Simulation experiments with two combustion timings were conducted to find
optimal u with respect to (10.10). The objective was to investigate when multiple
injections are beneficial and how optimal combustion timings are configured as
a function of imposed constraints. This section presents the model used, and
simulation results showing trends for p max , NOx , temperature at exhaust-valve
opening, TEVO , and η GIE as u is varied.

177
Chapter 10. Constraint Handling with Multiple Injections

Model
The model used was the zero-dimensional model presented in Chapter 2
µ ¶
dp γ dV γ − 1 dQ c dQ ht
=− p+ − , p(θIVC ) = p in (10.11)
dθ V dθ V dθ dθ

where γ was dependent on in-cylinder temperature and composition. The

heat-transfer rate dQ ht /d θ was computed with a Woschni-type convective
heat-transfer coefficient [Woschni, 1976], and the heat-release rate was modeled
with the Wiebe expression [Wiebe, 1970]

 ³ µθ−θ ¶b+1 ´
Q c (θ) 1 − exp −a SOC
for θ ≥ θSOC
= ∆θ (10.12)
Q ctot 
0 otherwise

i
Here, θSOC was related to the i :th injection timing θSOI with the use of an Arrhe-
nius ignition-delay expression
i i
θign = θSOI + τi
(10.13)
τi = Ap −n i e E a /R̃TSOIi
SOI

where ignition delay τi was omitted in (10.13) if combustion had started prior to
injection.
NOx formation was computed with the two-zone, Zeldovich-mechanism mo-
del presented in [Egnell, 2001]. A more detailed description of this model is given
in Chapter 2.
Instead of modeling Tex , it was decided to study the in-cylinder temperature
at exhaust-valve opening TEVO , which correlates with Tex .

Simulated Conditions
Simulation experiments were conducted with two injections and the correspon-
ding accumulated heat-release

Q c (θ) = r Q c1 (θ) + (1 − r )Q c2 (θ), r ∈ [0.5, 1] (10.14)

where Q c1 and Q c2 are Wiebe expressions on the form of (10.12) and r is defined by
(10.5). The combustion timings of Q c1 and Q c2 , θCT
1 2
and θCT were swept for diffe-
rent r and constant total fuel energy. The reason for limiting the study to two in-
jections and not considering a more general heat-release was that this would re-
semble a realistic scenario that could be realized in the experimental setup used.
Constraints on p max and NOx were evaluated at a higher load compared to TEVO ,
since lower bounds on TEVO are more likely to become active at low load. Model
parameters used are presented in Table. 10.1.

178
 10.3 Simulation

Table 10.1 Model parameters used in simulation. Constraints on p max and NOx
were evaluated at a higher load compared to TEVO .

p max and NOx constraints

Intake Conditions Wiebe Parameters

p in [bar] 0.98 Q ctot [J] 5 × 103

Tin [K] 293 a [-] 2.3
λ [-] 2.5 ∆θ [CAD] 6
r EGR [-] 0 b [-] 1.8

TEVO constraints

Intake Conditions Wiebe Parameters

p in [bar] 1.6 Q ctot [J] 3 × 103

Tin [K] 293 a [-] 2.3
λ [-] 2.5 ∆θ [CAD] 7.5
r EGR [-] 0 b [-] 1.8

All constraints

Cylinder Geometry Heat-Transfer

r c [-] 18 C 1 [-] 2.28

Vd [m3 ] 2.1 × 103 C 2 [m/(sK)] 0.0032
B [mm] 130 Tc [K] 333
L [mm] 160 m c c p [J/K] 1150
IVC [CAD] -151 k c [J/(mK)] 45
EVC [CAD] 146 L c [m] 0.025

Ignition Delay [Spadaccini and TeVelde, 1982]

A [CAD] 1.7496 × 10−8

n [-] 2
E a /R̃ [K] 20926

Simulation Results
1 2
Level curves for p max , NOx , TEVO and η GIE as a function of θCT and θCT are pre-
sented in Figs. 10.2, 10.3 and 10.4. Results with r = 0.5, 0.625, 0.75 and 0.875 are
presented in Fig 10.2, whereas results with r = 0.5 are presented in Figs. 10.3 and
10.4. In these figures, the black lines are η GIE level curves, with the most efficient
combustion timing, marked ×, found close to TDC. The efficiency then decreases

179
Chapter 10. Constraint Handling with Multiple Injections

1 2
with delayed θCT and θCT . The colored lines indicate p max [bar], NOx [ppm] and
TEVO [K] level curves, where the colored symbols indicate η GIE -optimal combus-
tion timings with respect to the different level curves as constraints, for both
1 2 1 2
θCT > θCT (∗) and θCT < θCT (¦). The gray regions indicate combustion-timings
where fuel-injection pulses overlap, i.e., u ∉ U.
Figures 10.2-10.4 show that it is more efficient with two combustion timings
when subject to p max , NOx and TEVO constraints. This is recognized by observing
1 2
that optimal timings occur away from the θCT = θCT line which indicate a unimo-
dal, single-injection heat release rate. This holds even if u ∉ U would have been
allowed.
1 2
The symmetry with respect to θCT = θCT in Figs. 10.3 and 10.4 is due to the
heat-release rates being identical with r = 0.5. The symmetry was altered as r
was varied in Fig. 10.2. In this figure, it can be seen that η GIE was maximized for
1 2 1 2
both θCT > θCT and θCT < θCT . In general, there was no trend favoring any of these
two configurations.
Trade-offs between η GIE and x c for different r are presented in Fig. 10.5. The
p max and NOx trade-offs can be improved by up to 4 % when using two combus-
tion timings instead of one (r = 1), whereas the TEVO trade-off can be improved
c
by up to 2 %. The potential η GIE advantage with two injections increases as p max
c c
and NOx become more conservative. This does not hold for TEVO . The trade-offs
in Fig. 10.5 do not distinguish any clear choice for r < 1.
An attempt to explain the observed trends is made in Fig. 10.6. This figure
presents efficiency-optimal combustion timings for one and two injections with
c
arbitrary p max (upper), NOcx (middle) and TEVO c
(lower). For p max and NOx con-
straints, two injections give a more distributed heat-release. This lowers the
peak pressure and temperature which gives a slower NOx -formation rate. Similar
trends were observed experimentally in [Han et al., 1996]. For the TEVO constraint
in the lower subdiagram, the late injection gives a sufficient contribution to TEVO
for the first combustion timing to be timed optimally. To summarize: the overall
trend in Fig. 10.6 is that two injections allow for an effective or mean combustion
timing closer to TDC, which increases the indicated efficiency η2GIE > η1GIE .

10.4 Controller Design

The optimization problem (10.10) was solved through simulation and by evalua-
ting a grid over u in the previous section. With the simulation results obtained,
the objective is to design a controller that automatically finds optimal u depen-
ding on operating conditions. The simulation results suggested that a single fuel
injection should be used to obtain the efficiency optimal point × when no con-
straints are active. A constraint is here said to be active if the constrained output
variable is equal to the constraint limit at optimal fuel injection.

180
 10.4 Controller Design

r = 0.875 r = 0.75

150

130

110

130
150

110
20 η = 0.513 20 η = 0.518

η = 0.527
θ 2CT [CAD]

θ 2CT [CAD]
η = 0.531
η = 0.538
η = 0.543
10 10

η = 0.546 η = 0.537 η = 0.538

η = 0.546
η = 0.528 η = 0.532
η = 0.517 η = 0.523
0 0
0 10 20 0 10 20
θ 1CT [CAD] θ 1CT [CAD]

r = 0.625 r = 0.5

130

110
130

110

20 20
η = 0.523 η = 0.526
θ 2CT [CAD]

θ 2CT [CAD]

η = 0.538
η = 0.537

10 η = 0.542 10 η = 0.541
150 150 η = 0.526

η = 0.546 η = 0.527 η = 0.546

η = 0.54 η = 0.541
η = 0.535 η = 0.538
0 0
0 10 20 0 10 20
θ 1CT [CAD] θ 1CT [CAD]

Figure 10.2 Level curves of η GIE (black) for different combustion timings θCT 1 ,
2 , and ratios r . Level curves for p
θCT max are presented in red (110 bar), blue (130
bar) and green (150 bar). The shaded gray areas correspond to infeasible injection
timings where the injection pulses overlap. For each r , the most efficient timings
are marked ×. The colored marks indicate the most efficient feasible points, given
1 > θ 2 and θ 1 <
the different p max constraints for the feasible regions where θCT CT CT
2 . The figure shows that two combustion timings are optimal when subject to
θCT
p max constraints. This is recognized by observing that optimal timings occur away
from the θCT1 = θ 2 line which indicate a unimodal, single-injection heat release
CT
rate. Pressure curves corresponding to the marked high-efficiency points for r =
0.5 are presented in Fig. 10.7.

181
Chapter 10. Constraint Handling with Multiple Injections

NOx level curves

40

30
θ 2CT [CAD]

20

10 0
0
10
1400 η = 0.518
η = 0.545 600
η = 0.535
η = 0.54
0
0 5 10 15 20 25 30 35 40
θ 1CT [CAD]

Figure 10.3 Simulated η GIE level curves (black) as a function of θCT1 and θ 2 .
CT
Level curves for NOx are presented in red (600 ppm), blue (1000 ppm) and green
(1400 ppm). The colored marks indicate η GIE -optimal combustion timings with
respect to the different NOx constraints. NOx -formation curves corresponding to
the marked high-efficiency points are presented in Fig. 10.8.

As constraints become active, additional injections should be introduced.

One injection should be kept close to TDC and the second should be delayed
in order to fulfill the constraint. The controller should also keep the constrained
output as close to the limit as possible. This behavior is more clearly illustrated
in Figs. 10.7, 10.8 and 10.9 where crank-angle resolved heat-release rates are pre-
sented together with p, NOx and T for the optimal timings in Figs. 10.2-10.4. The
dashed horizontal lines in Figs. 10.2-10.4 indicate x c .
c
It was decided to handle the different constraints separately: p max and NOcx
c
by introducing an early pilot injection and delaying both injections, and Tex by
introducing and adjusting a late post injection. This choice was motivated by the
trends in Fig. 10.9, where a late combustion timing was found to be optimal. The
system input was therefore redefined as
¡ 1 2 3
¢T
u = θSOI θSOI θSOI r1 r3 (10.15)

where r 1 and r 3 are given by

mx
rx = (10.16)
m x + m2

182
 10.4 Controller Design

Temperature level curves

η = 0.473
40

η = 0.498
30
θ 2CT [CAD]

η = 0.519
20 850
93
0

10

89
0
η = 0.538

0η
0 5 10 15 20 25 30 35 40
θ 1CT [CAD]

Figure 10.4 Simulated η GIE level curves as a function of θCT1 and θ 2 . Level cu-
CT
rves for TEVO are presented in red (930 K), blue (890 K) and green (850 K). It
is optimal to keep one injection close the optimal point ×, and then delay the
latter to fulfill the constraint. Temperature curves corresponding to the marked
high-efficiency points are presented in Fig. 10.9.

The indices 1, 2 and 3 represent the pilot, main and post injection, respectively.
The controller should also deliver the desired load output, which was handled by
adjusting the total fuel mass m tot
f
.
A simple controller design is attractive from an implementation perspective.
Therefore, a hybrid multiple-input multiple-output PI controller

u(k + 1) = u(k) + k p (e(k) − e(k − 1)) + k I (e(k − 1)) (10.17)

with gains k p and k I was designed to achieve the desired system behavior. The
suggested controller used combustion timings, peak pressure levels, the modeled
NOx -emission level and measured exhaust temperature as feedback signals. This
controller is presented in the remainder of this section.

Combustion Detection
When controlling multiple combustion timings, it is no longer sufficient to use
the crank-angle of 50 % burnt as feedback variable. This quantity might not be re-
lated to a physical combustion timing with a multimodal heat-release rate. The

183
Chapter 10. Constraint Handling with Multiple Injections

Trade-off Between NOcx & ηGIE c

Trade-off Between p max & ηGIE
0.55 0.55

0.5 0.5
ηGIE [-]

ηGIE [-]
r = 0.5 r = 0.5
0.45 0.45
r = 0.625 r = 0.625
r = 0.75 r = 0.75
r = 1.0 r =1
0.4 0.4
500 1,000 1,500 2,000 80 100 120 140
NOcx [ppm] cp max [bar]

c
Trade-off Between TEVO & ηGIE
0.55
r = 0.5
r = 0.625
r = 0.75
0.5 r = 1.0
ηGIE [-]

0.45

0.4
900 1,000
cTEVO [K]

Figure 10.5 Simulated x c / η GIE efficiency trade-offs for different r . Multiple

combustion timings can increase η GIE by up to 4 % with respect to NOcx and p max
c ,
c
and 2 % with respect to TEVO .

approach taken was to instead use the combustion-detection method presen-

x
ted in Chapter 4 to detect θCT from the pilot, main and post injections. In this
method, the heat-release rate dQ c /d θ is first computed. Then, the M most sig-
nificant peaks above a threshold level are detected, where M is the number of
injections used. The crank angles at the detected peaks then constitute the com-
bustion timings θCT . A minimum distance between detected peaks was introdu-
ced in order to not detect multiple peaks from one injection. This could otherwise
occur if a single-injection heat-release rate has multiple peaks, which is common
for conventional diesel combustion.

184
 10.4 Controller Design

200
One Injection
Two Injections
150 η1GIE = 0.476
p [bar]

100 η2GIE = 0.496

50

0
−20 −10 0 10 20 30 40 50
θ [CAD]

2,000
η1GIE = 0.49 p
NOx
1,500
η2GIE = 0.50
NOx [ppm]

1,000

500

0
−20 −10 0 10 20 30 40 50
θ [CAD]

2,000
η1GIE = 0.48 p
T
1,500 η2GIE = 0.49
T [K]

1,000

500

0
−20 0 20 40 60 80 100 120 140
θ [CAD]

Figure 10.6 A physical explanation for the difference between one (1) and two (2)
c
injections with arbitrary p max (upper), NOcx (middle) and TEVO
c (lower). For p max
and NOx constraints, two injections give a more distributed heat-release. A di-
stributed heat release lowers the peak pressure and gives a slower NOx -formation
rate, which allows for an effective or mean combustion timing closer to TDC. For
the TEVO constraint, the late injection provides a sufficient TEVO increase for the
first injection to be timed optimally.

185
Chapter 10. Constraint Handling with Multiple Injections

Optimal Combustion Timings

200

150
p [bar]

100

50

0
−20 −10 0 10 20 30 40 50
θ [CAD]

Figure 10.7 Efficiency-optimal combustion timings with respect to p max con-

straints and r = 0.5 in Fig. 10.2. When the constraint becomes increasingly strin-
gent, it is optimal to keep one combustion timing close to TDC and then delay the
latter. For a very stringent constraint the early combustion timing is also delayed.

Small fuel quantities are not always sufficient to generate heat-release. For in-
stance, in Chapter 8, a small pilot was only used to enhance the reactivity of the
main-injection and did not generate a separate heat-release impulse. The detec-
x
tion method therefore only allocated detected θCT to injections with sufficient
fuel mass. Values in the range of 15-30 mg were used as lower limits in the engine
experiments presented below. When a single combustion timing was expected,
θ50 was used instead of θCT as a combustion-timing indicator. A similar proce-
dure was applied for detecting the peak pressure levels generated by the pilot
1 2
and main injection p max , p max which are necessary quantities for the control-
c
ler to fulfill p max . The motored pressure curve was first subtracted from p before
1,2
detecting p max , to not detect the pressure peak generated by compression.

Without Active Constraints

2
Without active constraints, the main combustion timing θCT was set to follow
r 2
an efficient set point θCT through θSOI adjustment. The engine load p IMEPn was
controlled with m tot
f
r
for tracking of a set point p IMEPn . This controller behavior

186
 10.4 Controller Design

Optimal Combustion Timings

2,000

1,500
NOx [ppm]

1,000

500

−20 −10 0 10 20 30 40 50
θ [CAD]

Figure 10.8 Efficiency-optimal combustion timings with respect to the

NOx -emission constraints in Fig. 10.3. When NOcx becomes increasingly strin-
gent, it is optimal to keep the early combustion timing close to TDC and then
delay the latter. For a very restrictive constraint, the early combustion timing is
also delayed.

was obtained by the following cycle-to-cycle PI controller

2 2 θ θ
θSOI (k + 1) = θSOI (k) + k pCT (e θ2 (k) − e θ2 (k − 1)) + k I CT e θ2 (k − 1)
CT CT CT
p IMEPn (10.18)
m tot tot
f (k + 1) = m f (k) + k p (e p IMEPn (k) − e p IMEPn (k − 1))
p IMEPn
+ kI e p IMEPn (k − 1)

where k is cycle index, e p IMEPn and e θ2 are p IMEPn and θCT errors
CT

r
e θ2 (k) = θCT (k) − θθ2 (k)
CT CT
(10.19)
r
e p IMEPn (k) = p IMEPn (k) − p IMEPn (k)

Pressure Constraint
2 c
The main-injection timing θSOI was adjusted if p max was violated to keep the
2 c
pressure peak corresponding to the main injection, p max , below p max . This be-
havior was obtained by the following controller
( 2 2 c
θ (k) + ∆PICT (e θ2 (k)) if p max (k) < p max
θSOI (k + 1) = SOI
2
2
CT (10.20)
θSOI (k) + ∆PIp max (e p max
2 (k)) otherwise

187
Chapter 10. Constraint Handling with Multiple Injections

Optimal Combustion Timings

2,000

1,500
T [K]

1,000

500

−20 0 20 40 60 80 100 120 140

θ [CAD]

Figure 10.9 Efficiency-optimal combustion timings with respect to the

TEVO -emission constraints presented in Fig. 10.9. When the TEVO constraint
becomes increasingly stringent, it is optimal to keep the first injection timing
combustion close to TDC and then delay the latter.

where
2 c
e p max
2 (k) = p max (k) − p max (10.21)
and ∆PIx are combustion timing updates from two separate PI controllers on the
c
form of (10.17). A pilot injection was also introduced if p max > p max , in order to
obtain a distributed heat-release rate, see Fig. 10.7. This was done by varying r 1
according to
(
1 r 1 (k) + k r (r ∗1 − r 1 (k)) c
if p max (k) > p max
r (k + 1) = 1 1
(10.22)
r (k) − k r r (k) otherwise

where r ∗1 is a predefined set point. Furthermore, once the pilot injection was in-
1 c
troduced, a pilot-injection controller was used to keep p max below p max
( 1 1 c
1
θSOI (k) + ∆PICT (e θ1 (k)) if p max (k) < p max
θSOI (k + 1) = 1
CT (10.23)
θSOI (k) + ∆PIp max (e p max
1 (k)) otherwise

where
r 1
e θ1 (k) = θCT (k) − θCT (k)
CT
(10.24)
1 c
e p max
1 = p max − p max

188
 10.4 Controller Design

NOx Constraint
NOx -constraint fulfillment was handled similarly
(
2
2
θSOI (k) + ∆PICT (e θ2 (k)) if NOx (k) < NOcx
θSOI (k + 1) = CT (10.25)
2
θSOI (k) + ∆PINOx (e NOx (k)) otherwise

where
e NOx (k) = NOx (k) − NOcx (10.26)
A pilot injection was also introduced if NOx (k) > NOcx to obtain a distributed
heat-release rate according to Fig. 10.8
(
1 r 1 (k) + k r (r ∗1 − r 1 (k)) if NOx (k) > NOcx
r (k + 1) = 1 (10.27)
r (k) − k r r 1 (k) otherwise

Furthermore, as the pilot injection was introduced, the pilot-injection controller

(
1
1
θSOI (k) + ∆PICT (e θ1 (k)) if NOx (k) < NOcx
θSOI (k + 1) = CT (10.28)
1
θSOI (k) + ∆PINOx (e NOx (k)) otherwise

1
was used to adjust θSOI for constraint fulfillment.

Temperature Constraint
The exhaust temperature Tex was controlled by adjusting the post injection to
obtain the configurations presented in Fig 10.9
(
3 r 3 (k) + k r (r ∗3 − r 3 (k)) c
if Tex (k) < Tex
r (k + 1) = 3
r (k) − k r r 3 (k) otherwise
( (10.29)
3 c
3
θSOI (k) + ∆PICT (e θ3 (k)) if Tex (k) > Tex
θSOI (k + 1) = CT
3
θSOI (k) + ∆PITex (e Tex (k)) otherwise

where
r 3
e θ3 (k) = θCT (k) − θCT (k)
CT
(10.30)
c
e Tex = Tex − Tex

p IMEPn Control
The engine load was controlled by adjusting the total injected fuel mass using the
controller
m tot tot
f (k + 1) = m f (k) + ∆PIp IMEPn (e p IMEPn (k)) (10.31)

The commanded fuel was then distributed according m tot f

, r 1 and r 3 , and
fuel-injection durations were computed using the injector map in (10.6).

189
Chapter 10. Constraint Handling with Multiple Injections

Saturation
Saturation limits were introduced to enforce

u ∈U (10.32)

1 3
This was done by saturating the pilot and post injection timings θSOI , θSOI to
avoid overlap with the main injection. Unfortunately, this could lead to dead-
1 1,2
lock and constraint violation for p max (k) when regulating p max . A remedy to this
1 1 2
problem was to reduce r when p max (k) > p max (k)

r 1 (k + 1) = r 1 (k) − k r r 1 (k) if p max

1 2
(k) > p max (k) (10.33)

1,2,3
Additional minimum and maximum limits on θSOI , m tot
f
and r 1,3 were introdu-
ced as safety margins to ensure feasible injection timings and durations.

Slack Variables
Slack variables ²x were added to the constraint conditions above
c
p max (k) > p max − ²p max
NOx (k) > NOcx − ²NOx (10.34)
c
Tex (k) < Tex + ²Tex

to avoid limit cycles around the constraint limits.

Summary
The controllers presented above are summarized in Algorithm 4. Here, p max - and
NOx -constraint handling were merged using the max function. In this way, the
controller would adjust for the constraint demanding the latest θSOI , since both
p max and NOx decrease with θSOI . The notation

x̃ c = x c ± ²x (10.35)

was introduced in Algorithm 4 to ease notation.

10.5 Experimental Evaluation

The controller in Algorithm 4 was evaluated experimentally, both in transient and
steady-state operation. The different constraints were investigated separately in
both cases. The controller was also compared to the single-injection controller
presented in Algorithm 5. This controller delayed a single injection timing to ful-
fill the constraints. Controller parameters used are presented in Table 10.2, where
the parameters were tuned in favor of robustness over convergence rate.

190
 10.5 Experimental Evaluation

Algorithm 4 Merged Constraint Controller

Pilot-Injection Controller
c ˜ x then c
1: if p max (k) > p̃ max ∨ NOx (k) > NO
1 1
¡ ¢
2: θSOI (k + 1) = θSOI (k) + max ∆PIp max (e p max
1 (k)), ∆PINOx (e NOx (k))
c 1 2
3: if p max (k) > p̃ max ∧ p max (k) > p max (k) then
4: r 1 (k + 1) = r 1 (k) − k r r 1 (k)
5: else
6: r 1 (k + 1) = r 1 (k) + k r (r ∗1 − r 1 (k))
7: end if
8: else
1 1
9: θSOI (k + 1) = θSOI (k) + ∆PICT (e θ1 (k))
CT
1 1 1
10: r (k + 1) = r (k) − k r r (k)
11: end if
Main-Injection Controller
c ˜ x then c
12: if p max (k) > p̃ max ∨ NOx (k) > NO
2 2
¡ ¢
13: θSOI (k + 1) = θSOI (k) + max ∆PIp max (e p max
2 (k)), ∆PINOx (e NOx (k))
14: else
2 2
15: θSOI (k + 1) = θSOI (k) + ∆PICT (e θ2 (k))
CT

16: end if
Post-Injection Controller
17: if Tex (k) < T̃ex then
3 3
18: θSOI (k + 1) = θSOI (k) + ∆PITex (e Tex (k))
19: r 3 (k + 1) = r 3 (k) + k r (r ∗3 − r 3 (k))
20: else
3 3
21: θSOI (k + 1) = θSOI (k) + ∆PICT (e θ3 (k))
CT

22: r 3 (k + 1) = r 3 (k) − k r r 3 (k)

23: end if

191
Chapter 10. Constraint Handling with Multiple Injections

Algorithm 5 Single Injection Controller

Main-Injection Controller
c c
˜ x ∨ Tex (k) < T̃ex
c
1: if p max (k) > p̃ max ∨ NOx (k) > NO then
³ ´
2
2: ∆θSOI (k) = max ∆PIp max (e p max
2 (k)), ∆PINOx (e NOx (k)), ∆PITex (e Tex (k))
2 2 2
3: θSOI (k + 1) = θSOI (k) + ∆θSOI (k)
4: else
2 2
5: θSOI (k + 1) = θSOI (k) + ∆PICT (e θ2 (k))
CT

6: end if

Transient Operation
c
p max c
The ability to handle p max constraints was evaluated with p max = 75. Ex-
r
perimental results with p IMEPn step changes are presented in Fig. 10.10, where
r
p IMEPn increased from 4 to 10 bar in 10 cycles as p IMEPn was increased.
2 r c c
At p IMEPn = 10 bar with θCT = θCT , p max violated p max . To fulfill p max , the con-
1 2
troller increased r and delayed θCT , and reached a new injection configuration
r
in 50 cycles. The controller converged to the initial conditions once p IMEPn was
decreased.
Cycle-resolved data at cycles 1250 and 1450 are presented in Fig. 10.11. At
2 r
cycle 1250, a single injection was used where p IMEPn = 4 bar, and θCT = θCT =8
1
CAD. At cycle 1450, with r = 0.5, the injection timings were adjusted to fulfill
c 1 2 r
p max . The red and blue vertical lines indicate detected θCT and θCT , and θCT is
indicated by the horizontal dashed line.
NOx c Fulfillment of NOcx = 800 ppm is presented in Fig. 10.12. In this figure,
the controller was compared to the single-injection controller in Algorithm 5,
which is indicated in purple. The combustion timings were delayed when NOx
r
was increased due to positive p IMEPn step changes. The double-injection control-
ler introduced a pilot injection when NOcx was violated, and an additional com-

Table 10.2 Controller parameters used in experiment.

θ θ
k pCT [-] 0.1 k I CT [-] 0.15
NO NO
kp x [CAD/ppm] 10 k I x [CAD/ppm] 15
p p
k p max [CAD/bar] 0.04 k I max [CAD/bar] 0.05
T T
k p ex [CAD/K] 0.3 k I ex [CAD/K] 0.35
mf mf
kp [mg/bar] 1.5 kI [mg/bar] 2

192
 10.5 Experimental Evaluation

Load Tracking Maximum Cylinder Pressure

15 100
r
p IMEPn p IMEPn 1 2 c
p IMEPn [bar] p max p max p max

10

p max [bar]
80

5
60

0
1,200 1,300 1,400 1,500 1,600 1,200 1,300 1,400 1,500 1,600
cycle [-] cycle [-]

Injection and Combustion Timings Pilot Ratio

20 1

0.8
θ SOI / θ CT [CAD]

10
0.6
r 1 [-]

0.4
0
0.2

−10 0
1,200 1,300 1,400 1,500 1,600 1,200 1,300 1,400 1,500 1,600
cycle [-] cycle [-]

θ 1CT θ 2CT θ 1SOI θ 2SOI θ rCT

Figure 10.10 Experimental results showing p max constraint handling. As p IMEPnr

c
was increased from 4 to 10 bar, p max with θ50 = 8 violated p max . Constraint viola-
tion forced r 1 to increase and θCT
2 to be delayed. The controller converged to the
r
initial conditions once p IMEPn was decreased.

bustion timing (red) was detected. The fuel-masses presented are the fuel masses
demanded by the controllers, and not the actual injected fuel mass. Both con-
trollers had comparable p IMEPn response times of 10 cycles and injection-timing
settling times of 25 cycles. The double-injection controller had a slightly larger
NOx overshoot.
c
Tex c
Fulfillment of Tex = 240◦ C is presented in Fig. 10.13. The combustion ti-
ming was delayed for the single-injection controller in order to fulfill the con-
r
straint when Tex was decreased due to negative p IMEPn step changes. A post in-

193
Chapter 10. Constraint Handling with Multiple Injections

Cycle: 1250 Cycle: 1450

100
p [bar] 100

p [bar]
50 50

0 0
0 20 0 20
θ [CAD] θ [CAD]

Figure 10.11 Cycle-resolved data belonging to cycle 1250 and 1450 in Fig. 10.10.
1,2 r . The dashed horizontal lines indicate
The vertical lines indicate θCT and θCT
c
p max .

jection was introduced by the double-injection controller and two combustion

2
timings were detected. The Tex increase caused by the post injection allowed θCT
r
to follow θCT . The Tex loop was considerably slower than the other loops due to
heat-transfer dynamics in the exhaust system with Tex settling times of approxi-
mately 100-200 cycles.

Steady-State Operation and Trade-offs

The controller-design effect on emission and efficiency trade-offs was evaluated
c
in steady-state. An efficiency improvement with respect to p max = 95 bar is pre-
sented in Fig. 10.14 where the constraint was fulfilled with one (blue) and two
(red) injections. Two injections allowed for a combustion timing closer to TDC
which increased η GIE with 5 %. A drawback with the double-injection strategy
was the significant increase in soot emissions.
More detailed NOcx and Texc
trade-offs with respect to η NIE , NOx , soot and HC
emissions are presented in Figs. 10.15 and 10.16. Pressure and heat-release rates
for the different x c sweeps are presented in Fig. 10.17. The results in Fig. 10.15
show that the suggested controller improved the NOx trade-offs with η NIE and
HC, and worsened the trade off with respect to soot and also partly with Tex . Fi-
gure 10.16 shows that the suggested controller partly worsened the Tex trade-off
c
with respect to η NIE and NOx for high Tex , but improved the HC trade-off. The
worsened trade-off with η NIE contradicts the simulation results presented in
Fig. 10.4.
The overall soot increase with two injections is believed to be caused by shor-
tened mixing times and fuel injection during combustion. The improved HC

194
 10.5 Experimental Evaluation

1,500
Two Injections
NOx [ppm]
1,000

500
One Injection
0
1,000 1,100 1,200 1,300 1,400 1,500 1,600 1,700 1,800 1,900 2,000
cycle [-]

p IMEPn S r
p IMEPn r,S
10 p IMEPn p IMEPn
p IMEPn [bar]

4
1,000 1,100 1,200 1,300 1,400 1,500 1,600 1,700 1,800 1,900 2,000
cycle [-]
m 1f m 2f m Sf

100
m f [mg]

50

0
1,000 1,100 1,200 1,300 1,400 1,500 1,600 1,700 1,800 1,900 2,000
cycle [-]

θ 1SOI θ 2SOI θ 1CT θ 2CT θ SSOI θ SCT θ rCT r,S

θ CT

20
θ SOI / θ CT [CAD]

−20
1,000 1,100 1,200 1,300 1,400 1,500 1,600 1,700 1,800 1,900 2,000
cycle [-]

Figure 10.12 A transient NOx -control experiment with a comparison between

the single- (purple, S) and double-injection controller. The combustion timings
r
were delayed when NOx increased due to positive p IMEPn step changes. A pilot
injection was introduced (red) and two combustion timings were detected by the
double-injection controller.

195
Chapter 10. Constraint Handling with Multiple Injections

320
Tex S
Tex c Tex
300 One Injection Two Injections
Tex [◦ C] 280
260
240
220
200 400 600 800 1,000 1,200 1,400 1,600 1,800 2,000
cycle [-]

12 S r r,S
p IMEPn p IMEPn p IMEPn p IMEPn
10
p IMEPn [bar]

8
6
4
200 400 600 800 1,000 1,200 1,400 1,600 1,800 2,000
cycle [-]

m 2f m 3f m Sf
100
m f [mg]

50

0
200 400 600 800 1,000 1,200 1,400 1,600 1,800 2,000
cycle [-]

θ 2SOI θ 3SOI θ 2CT θ 3CT θ SSOI θ SCT

20
θ SOI / θ CT [CAD]

−20
200 400 600 800 1,000 1,200 1,400 1,600 1,800 2,000
cycle [-]

Figure 10.13 A transient Tex -controller experiment with a comparison between

the single- (purple) and double-injection controller. A post injection was introdu-
ced (green) and two combustion timings were detected by the double-injection
2 to follow θ r .
controller. This enabled θCT CT

196
 10.6 Discussion

Pressure Constraint
120
ηGIE = 0.442 [-] ηGIE = 0.468 [-]
100 NOx = 2111 [ppm] NOx = 1860 [ppm]

Soot = 4.1 [mg/m3 ] Soot = 13.4 [mg/m3 ]

80 HC = 13.2 [ppm] HC = 14.8 [ppm]
bsfc = 218 [g/kWh] bsfc = 208 [g/kWh]
p [bar]

60

40

20

−40 −30 −20 −10 0 10 20 30 40 50 60

θ [CAD]

c
Figure 10.14 Fulfillment of p max = 95 bar with one (blue) and two (red)
injections. Two injections allowed for a combustion timing closer to TDC,
which increased gross-indicated efficiency with 5 %. A disadvantage with the
double-injection strategy was the significant increase in soot emissions.

trade-offs with two injection could be explained by the advanced combustion

timing. It has also been shown that post injections can be used to aid oxidation
of HC [Chartier et al., 2011].

10.6 Discussion
From the experimental results in Sec. 10.5, it can be concluded that multiple
injections can increase the indicated efficiency when stringent constraints on
p max and NOx are imposed. The efficiency increase was a result of the distribu-
ted heat-release rate with reduced peak in-cylinder pressure and temperature,
which allowed for a more advanced effective combustion timing, see Fig. 10.6.
These results agree with previous simulation and experimental work on optimal
heat-release rates [Eriksson and Sivertsson, 2016; Okamoto and Uchida, 2016;
Guardiola et al., 2017].
The problem of optimally calibrating multiple injection timings and dura-
tions for different engine operating points is both demanding and sensitive
to disturbances. The contribution of the work presented in this chapter is a

197
Chapter 10. Constraint Handling with Multiple Injections

0.49 5

0.48 4.5

Soot [mg/m3 ]
η NIE [-]

0.47 4

0.46 3.5
two injections
one injection
0.45 3
400 600 800 1,000 1,200 400 600 800 1,000 1,200
NOx [ppm] NOx [ppm]

45 340

40 320
HC [ppm]

35 Tex [◦ C] 300

30 280

25 260
400 600 800 1,000 1,200 400 600 800 1,000 1,200
NOx [ppm] NOx [ppm]

Figure 10.15 Steady-state NOx trade-offs with the single and double-injection
controller. The double-injection controller improved the NOx trade-off with η NIE
by 4 %. It also improved the trade-off with HC. The double-injection controller
had a worsened trade off with respect to soot, and a partly worsened trade-off
with Tex .

feedback controller that automatically sets the number of injections, timings

and durations with pressure-sensor measurements. This allows for operation
close to the constraint limits, which increases efficiency. This work differs from
previous research, which have, to the author’s knowledge, mainly discussed op-
timal heat-release rates in open loop.
The controller design was motivated by simulation results obtained from a
0D model with a bimodal heat-release rate, which showed a 2-4 % efficiency in-
crease with respect to constraints on p, NOx and TEVO . The model was mainly
chosen for its simplicity. This study should be redone with respect to more detai-
led fuel-injection, mixing and combustion models to determine if the observed
trends still hold when additional effects are considered. This study could also be
extended to cover a larger engine operating range, to evaluate how constraint

198
 10.6 Discussion

0.48

Soot [mg/m3 ]
0.47
η NIE [-]

0.46
two injections
2
one injection
0.45
210 220 230 240 250 210 220 230 240 250
Tex [◦ C] Tex [◦ C]

200 800

150 700

NOx [ppm]
HC [ppm]

100 600

50 500

0 400
210 220 230 240 250 210 220 230 240 250
Tex [◦ C] Tex [◦ C]

Figure 10.16 Steady-state Tex trade-offs with the single and double-injection
controller. The results show that the double-injection controller worsened the Tex
trade-off with η NIE and NOx for high Tex c , but improved the trade-off with HC

emissions.

trade-offs vary with engine load and speed.

A hybrid PI controller was designed with the objective of finding the optimal
injection configurations computed in simulation. The controller design was mo-
tivated by its simplicity, where tuning parameters were PI controller gains and
constraint thresholds for varying the number of injections. A drawback with a
hybrid controller design of this kind is however the difficulty of finding condi-
tions for stability, which in this work, was obtained through manual tuning. Con-
straint limits were chosen arbitrarily in the experimental evaluation to demons-
trate that the desired controller behavior was achieved. Suggested future work is
to set these limits according to actual engine constraints.
The experimental controller evaluation showed injection-timing settling ti-
mes of 25-50 cycles with respect to constraints on p max and NOx during p IMEPn
changes of 10 cycles. Injection-timing settling times during Tex constraint ful-

199
Chapter 10. Constraint Handling with Multiple Injections

NOcx Sweep, One Injection NOcx Sweep, Two Injections

p [bar] 100 100

p [bar]
50 50

0 0
0 20 40 0 20 40
θ [CAD] θ [CAD]

c Sweep, One Injection

Tex c Sweep, Two Injections
Tex
100 100
p [bar]

p [bar]

50 50

0 0
0 20 40 0 20 40
θ [CAD] θ [CAD]

Figure 10.17 Pressure and heat-release rates for the constraint sweeps presented
in Figs. 10.15 and 10.16.

fillment were 100-200 cycles. These settling times were comparable to those of
the single-injection controller, that adjusted one injection timing to fulfill con-
straints.
Steady-state experiments showed a 4-5 % efficiency increase with respect to
p max and NOx constraints, and a 1.7 % efficiency decrease with respect to Tex
c
constraints. The decrease in efficiency with Tex contradicted the simulation re-
sults in Sec. 10.3. One possible explanation for the efficiency decrease could be
the assumption of a constant heat-release shape with combustion timing. In re-
ality, heat-release rates and combustion efficiency decrease with combustion ti-
ming, which penalize the efficiency of late post injections. It was however not
easy to verify this from the heat-release rates in Fig. 10.17. Despite this fact, the
controller suggested provides a framework for how to introduce post injections
when favorable. Results presented in [Honardar et al., 2011] showed that a post

200
 10.7 Conclusions

injection could provide a small fuel-consumption reduction compared to that of

a delayed main injection when increasing the exhaust temperature.
Although the controller design was motivated for its simplicity, the control-
ler complexity became rather involved, see Algorithm 4. An alternative approach
would be to design a model predictive controller that solves (10.10) on-line. The
MPC framework could then provide an increased flexibility when adding and re-
moving constraints and injections in (10.10). Such a controller design is conside-
red in Chapter 12.

10.7 Conclusions
A controller was designed and implemented for increased thermodynamic effi-
ciency when constraints on p max , NOx and Tex are imposed. The controller de-
sign was motivated by previous research on multiple injections and a presented
0D-simulation study that showed a 2-4 % efficiency increase when introducing
an additional injection. These results suggested that the controller should ad-
just the number of injections for efficient constraint fulfillment. This was done
by introducing a pilot injection when encountering active p max and NOx con-
straints, whilst exhaust-temperature constraints were handled with a post injec-
tion. A single injection was found optimal when no constraints were active.
The desired controller behavior was obtained with a hybrid, multiple-input
multiple-output PI controller that utilized feedback from in-cylinder pressure
measurements, a NOx -emission model that functioned as a virtual sensor, and
measured exhaust temperature. The suggested controller was experimentally
evaluated, where it showed comparable transient performance to that of a
single-injection controller. The controller exhibited improved p max / η NIE and
NOx / η NIE trade-offs with a 4-5 % increase in η NIE . The controller also showed a
worsened Tex / η NIE trade-off with a 1.7 % decrease in η NIE . Increased soot emis-
sions levels with two injections were also observed.

201
11
Pressure Prediction and
Efficiency Optimization

11.1 Introduction
Combustion timing has traditionally been controlled in open loop by means of
experimentally calibrated injection-timing maps [Guzella and Onder, 2009]. This
approach requires a considerable calibration effort and can be sensitive to va-
riations in hardware and fuel properties, especially in low-temperature combus-
tion modes. Experimental results in the previous chapters of this thesis have
shown that closed-loop combustion-timing control can be used to accurately
track combustion-timing set points and make the combustion timing robust to
disturbances.
Closed-loop combustion-timing controllers can be divided into two sub-
groups, one where the controller tracks a predefined combustion-timing set
point, where set-point optimization is considered to be a separate task. Exam-
ples of such controllers were discussed in chapters 6, 7 and 9, where PI control
and MPC were used for set-point tracking. Then, there are controllers that in-
stead adjust the combustion timing to directly fulfill higher-level performance
targets, such as emission-limit fulfillment and efficiency maximization. An ex-
ample of a high-level performance controller was presented in [Karlsson et al.,
2010], where a dynamical black-box model related the injection timing to p IMEPg ,
d p max and NOx emissions to minimize fuel consumption subject to specified
output constraints. Similar data-driven approaches were presented in [Hafner
et al., 2000; Atkinson et al., 2009]. Extremum-seeking control is another example
of a data-driven controller design that aims to fulfill higher-level specifications.
Extremum-seeking control has previously been used to find efficiency-optimal
combustion timings through set-point perturbation, see [Lewander et al. 2012].
This technique was also investigated by Killingsworth et al. [2009] and Hellström
et al. [2013], in HCCI- and spark-ignition engines, respectively.

202
 11.2 Controller Description

A desirable controller feature would be to utilize physical knowledge in the

combustion-timing optimization. However, the model complexity needed to de-
scribe the steps from fuel injection to heat-release and cylinder pressure makes
the design and implementation of such a controller difficult. Approximate mo-
deling approaches are therefore needed to simplify the controller design.
This chapter introduces a physics-based cylinder-pressure controller that
takes advantage of the estimated heat-release rate to predict how the cylinder
pressure varies with fuel-injection timing. This method allows for high-level
combustion-timing control, and keeps modeling complexity at a manageable
level. Predicted pressure variation is computed with a linearized 0D model,
where the linearization is conducted with respect to the previous engine-cycle
pressure and heat-release rate. The predicted cylinder-pressure variation is
then used by the controller to maximize indicated efficiency without violating
cylinder-pressure constraints.
The controller algorithm is described in Sec. 11.2. The prediction method and
closed-loop performance is then experimentally evaluated in Sec. 11.3, together
with a parameter-sensitivity analysis. Discussion and conclusions are given in
Secs. 11.4 and 11.5.

11.2 Controller Description

The cycle-to-cycle controller can be summarized in three steps:

1. First, the heat-release rate dQ c /d θ is estimated from the previous-cycle

pressure signal using (4.1).

2. Then, the predicted pressure variation with respect to a crank-angle shift

∆θ in dQ c /d θ is computed. The pressure variation is computed by lineari-
zing the cylinder-pressure model in (2.9), with respect to the previous-cycle
cylinder pressure and dQ c /d θ.

3. Finally, the desired shift in injection timing ∆θSOI is obtained by solving an

optimization problem. The optimization problem is based on the predic-
ted cylinder pressure obtained in step 2 and aims to optimize the indicated
efficiency η GIE , subject to constraints on maximum cylinder pressure p max
and pressure-rise rate d p max .

Steps 1 to 3 are explained in the following sections.

Pressure Prediction
First, the controller obtains the previous-cycle cylinder pressure p 0 and compu-
tes the corresponding heat-release rate dQ c0 /d θ using (4.1).
After computing dQ c0 /d θ, the objective is to predict how the pressure chan-
ges with combustion timing. A pressure change due to a change in combustion

203
Chapter 11. Pressure Prediction and Efficiency Optimization

timing was here assumed to be equivalent to the effect of shifting dQ c0 /d θ as a

function of θ. This assumption relies on weak cycle-to-cycle dynamics and small
dQ c0 /d θ-shape variations with smaller changes in combustion timing. The com-
puted dQ c0 /d θ is therefore shifted a crank angle ∆θ

dQ c+ dQ c0 (θ + ∆θ)
=
dθ dθ
(11.1)
dQ c− dQ c0 (θ − ∆θ)
=
dθ dθ

In-cylinder pressure curves p + and p − , corresponding to dQ c+ /d θ and

dQ c− /d θ are then computed by linearizing the model
µ µ ¶¶
dp γ dV γ − 1 dQ c pV TIVC
=− p+ − hc A − Tw (11.2)
dθ V dθ V dθ p IVC VIVC

with respect to p at the previous-cycle pressure p 0 . Constant γ and T w were as-

sumed to simplify computations. The linearized pressure dynamics are given by
µ ¶
d ∆p γ dV ∂µ(p 0 , θ) γ − 1 d ∆Q c
=− + ∆p + (11.3)
dθ V dθ ∂p V dθ

where ∆p is the first order deviation from p 0 , d ∆Q c /d θ is the deviation from

dQ c0 /d θ, and the nonlinear term in the heat-transfer model is denoted

h c ATIVC
µ(p, θ) = (γ − 1) p (11.4)
p IVC VIVC

The convective heat-transfer coefficient used here is given by

h c = αB 0.2 p 0.8 T −0.55 ω0.8 (11.5)

(see, (2.15)), where α was introduced as a heat-transfer tuning parameter. The re-
ason for linearizing (11.2) is that ∆p can be computed from the solution to (11.3)
Z θ d ∆Q c (ϑ)
∆p(θ) = Φ(θ, ϑ)Γ(ϑ) dϑ (11.6)
θIVC dϑ

where µ Z ¶µ ¶γ
θ ∂µ(p 0 , τ) V (ϑ)
Φ(θ, ϑ) = exp − dτ
ϑ ∂p V (θ)
(11.7)
γ−1
Γ(ϑ) =
V (ϑ)

204
 11.2 Controller Description

The pressure deviation ∆p is therefore computationally cheap to generate, which

allows for on-line computations. Another motivation for the linearization is the
linear relation between d ∆Q c /d θ and ∆p, which is suitable for optimization and
linear MPC. The pressure trajectories p + and p − are given by p 0 + ∆p, where ∆p
is the solution to (11.3) with the inputs

d ∆Q c+ dQ c+ dQ c0
= −
dθ dθ dθ
(11.8)
d ∆Q c− dQ c− dQ c0
= −
dθ dθ dθ

Injection-Timing Optimization
With p + and p − , variations in quantities such as p IMEPg , p max and d p max can be
computed

Z VEVO Z VEVO
+ 1 1
p IMEPg = p + dV −
p IMEPg = p − dV
Vd VIVC Vd VIVC

+
p max = maxp + −
p max = maxp − (11.9)
θ θ

+
d p max = maxd p + −
d p max = maxd p −
θ θ

Furthermore, approximate numerical derivatives of these quantities with respect

to ∆θ are given by
+ −
∂p IMEPg p IMEPg − p IMEPg
≈
∂∆θ 2∆θ
+ −
∂p max p max − p max (11.10)
≈
∂∆θ 2∆θ
+ −
∂(d p max ) d p max − d p max
≈
∂∆θ 2∆θ

A simple model for p IMEPg , p max and d p max in the subsequent engine cycle can
then be formulated with the partial derivatives in (11.10)

 
∂p IMEPg
   
p IMEPg p0  ∂∆θ 
 
   IMEPg   
 p max  =  p 0  +  ∂p max  ∆θ (11.11)
   max   
0
 ∂∆θ 
d p max d p max  ∂(d p ) 
max
∂∆θ

205
Chapter 11. Pressure Prediction and Efficiency Optimization

An optimization problem in ∆θ was then constructed with the objective of maxi-

mizing work output whilst fulfilling constraints on p max and d p max

minimize − p IMEPg + β(∆θ)2 (11.12)

∆θ
 
∂p IMEPg
   0   ∂∆θ 
p IMEPg p IMEPg  
 
subject to  p max  =  p 0  +  ∂p max  ∆θ
max  
d p max 0
d p max  ∂∆θ 
 ∂(d p ) 
max
∂∆θ
   c 
p max p max
c
d p max  ≤ d p max 
c
|∆θ| ∆θ

Here, β is a positive cost weight that penalizes combustion-timing changes, and

c c
p max , d p max and ∆θ c are upper limits on p max , d p max and the absolute value of
∆θ.
The solution to (11.12) without constraints is given by

∗ 1 d p IMEPg
∆θuc = (11.13)
2β d ∆θ
c c
Combustion-timing shifts ∆θ that reach the constraint limits p max and d p max are
given by
p c − p max
0
∆θp max
c = max
∂p max /∂∆θ
(11.14)
c 0
d p max − d p max
∆θd p max
c =
∂(d p max )/∂∆θ
Now, if one assumes that p max and d p max are monotonically decreasing with ∆θ,
the solution to (11.12), ∆θ ∗ , is simply given by the largest value among ∆θuc∗
,
c
∆θp max and ∆θd p max , saturated within the limits of ∆θ . Furthermore, the opti-
c c

mization problem can be rephrased as a problem in injection timing ∆θSOI if

∂∆θ/∂θSOI is known. This partial derivative was here assumed to be equal to 1
for simplicity.
The steps of the cycle-to-cycle controller have now been defined and are
summarized in Algorithm 6 where k denotes cycle index.

11.3 Results
This section presents experimental controller results. Open-loop experiments
are first presented for the purpose of evaluating how well the controller predicts

206
 11.3 Results

Algorithm 6 Cylinder-Pressure Controller

1: while true do

2: Estimate dQ c0 /d θ with the measured p 0 at cycle k and (4.1)

3: Shift dQ c0 /d θ according to (11.1)
4: Compute p + and p − with (11.6)
5: Compute the partial derivatives in (11.10), and solve (11.12)
6: Set θSOI (k + 1) = θSOI (k) + ∆θ ∗ (k)
7: end while

changes in cylinder pressure with ∆θSOI . Closed-loop performance is then pre-

sented, where convergence, parameter sensitivity and constraint fulfillment are
evaluated.

Open-loop Experiments
The injection timing θSOI was swept with a single injection for three different
speed/load combinations to evaluate the pressure-prediction method. Condi-
tions at these three operating points are presented in Table 11.1. Each sweep
consisted of 2000 cycles where θSOI was incremented in steps of one from ap-
proximately -25 to 5 CAD after TDC.

Table 11.1 The data used in the prediction- and controller evaluation were ob-
tained from the following operating points.

Operating Point 1 2 3

Nspeed [rpm] 1200 1200 1500

θDOI [ms] 1.0 1.6 1.0
p rail [bar] 800 800 800
p in [bar] 1.0 1.5 1.2
Tin [◦ C] 15 45 35
λ [-] 2.5 1.5 2.2
r EGR [-] 0 0 0

207
Chapter 11. Pressure Prediction and Efficiency Optimization

Operating Point 1
0.52
0.5
ηGIE [-]
0.48
0.46
0.44
0.42
−10 −5 0 5 10 15 20
θ 50 [CAD]
Operating Point 2
0.52
0.5
ηGIE [-]

0.48
0.46
0.44
0.42
−10 −5 0 5 10 15 20
θ 50 [CAD]
Operating Point 3
0.52
0.5
ηGIE [-]

0.48
0.46
0.44
0.42
−10 −5 0 5 10 15 20
θ 50 [CAD]

Figure 11.1 Gross indicated efficiency η GIE as a function of θ50 at operating

points 1-3. The ◦-markers are sampled η GIE data, and the solid red line indica-
tes the estimated η GIE sample mean as a function of θ50 . The found most efficient
θ50 are indicated by ¦.

Indicated Efficiency Figure 11.1 shows the gross indicated efficiency

p IMEPg Vd
η GIE = (11.15)
m f Q LHV
as a function of θ50 for the three operating points. In Fig. 11.1, the ◦-markers are
sampled η GIE data, and the solid red line is the estimated η GIE sample mean as a
function of θ50 . The spread in η GIE was due to cycle-to-cycle variation.
Figure
£ ¤11.1 shows that the measured η GIE had a shallow maximum in
θ50 ∈ 0, 5 CAD for all operating points. The indicated efficiency then decreased

208
 11.3 Results

20
OP1 OP2 OP3

15

10
θ 50 [CAD]

−5
−20 −15 −10 −5 0
θ SOI [CAD]

Figure 11.2 The relation between θ50 and θSOI for the θSOI sweeps in Fig. 11.1

as θ50 was increased or decreased outside this interval. The efficiency was sli-
ghtly higher at operating point 1. This is believed to be a result of the difference
in intake temperature Tin , which was due to engine warm up and the order of
which the experiments were conducted.

∂θ50 /∂θSOI Figure 11.2 shows θ50 as a function of θSOI for the three θSOI sweeps
in Fig. 11.1. The assumption of a constant partial derivative ∂θ50 /∂θSOI was ac-
curate close to θ50 = 5. However, at the lower-load operating points, ∂θ50 /∂θSOI
decreased at early θSOI and increased at late θSOI . This can be explained by the in-
crease in ignition delay τ when θSOI was decreased or increased. This effect was
then stronger at low load where τ was longer. Similar trends for ∂θ50 /∂θSOI with
engine load was observed in Chapter 6 and Fig. 6.3.

Pressure Prediction The pressure-prediction performance was evaluated by

comparing how well the predicted pressure agreed with measured pressure data.
The black solid pressure curve in Fig. 11.3 is the cycle-averaged cylinder pres-
sure p 0 for an arbitrary injection timing θSOI
0
. The red and blue pressure curves to
the left and right of this curve are measured pressure curves p − and p + for θSOI
0
shifted ±1 CAD relative to θSOI . The dashed red and blue curves are the predicted
pressures p̂ − and p̂ + , computed with (11.6) and the same θSOI shifts. It can be
seen that the predicted pressure curves p̂ − and p̂ + agree fairly well with p − and
p +.
Figure 11.4 compares the pressure differences p + −p 0 and p − −p 0 in Fig. 11.3
with the prediction errors p + − p̂ + (dashed, blue) and p − − p̂ − (dashed, red). The
relative error increased with θ, which indicates that the model was unable to pre-
dict pressure changes related to the heat release during end of combustion. Inte-

209
Chapter 11. Pressure Prediction and Efficiency Optimization

140
p0
p−
120
p+
p̂ −
100 p̂ +
p [bar]

80

60

40
−20 −10 0 10 20 30
θ [CAD]

Figure 11.3 The black solid pressure curve is the cycle-averaged pressure p 0 for
0 . The red and blue pressure curves to the left and right of this curve are the
θSOI
0 . The
cycle-averaged pressures p + and p − for θSOI shifted ±1 CAD relative to θSOI
dashed red and blue curves are the predicted pressures p̂ − and p̂ + , computed
with (11.6) and the same θSOI shifts.

restingly, it can also be seen that the pressure deviation resulting from a θSOI shift
closely resembled a heat-release rate.
Figures 11.5-11.7 presents cycle-averaged pressure changes p ± − p 0 toge-
ther with the prediction errors p ± − p̂ ± (as in Fig. 11.4) for all θSOI at operating
points 1-3. Just as in Fig. 11.4, the blue (red) color indicate a pressure change
p + − p 0 (p − − p 0 ) due to a delayed (advanced) θSOI . It can be seen that the pre-
diction error was smaller for θSOI close to TDC for all operating points and that
the error gradually increased when θSOI was advanced or delayed. The prediction
error also changed sign somewhere around TDC. For the red lines, indicating a
delayed θSOI , this meant that p̂ − < p − for late θSOI , and p̂ − > p − for early θSOI .
The opposite trend was found for p̂ + and p + .
The pressure-prediction performance was also evaluated by computing the
coefficient of determination R 2

||p + − p̂ + ||22 + ||p − − p̂ − ||22

R2 = 1 − (11.16)
||p + − p 0 ||22 + ||p − − p 0 ||22

which is a common model-evaluation statistic. The R 2 score describes the frac-

tion of variance in the data explained by the model [Casella and Berger, 2002].
Figure 11.8 presents the R 2 score as a function of θ50 for the three θSOI sweeps
in Figs. 11.5-11.7. The pressure-prediction method worked well for θ50 ∈ [0, 6]
where R 2 ≥ 0.9. The performance then started to degrade outside this interval,

210
 11.3 Results

Pressure Deviation and Prediction Error

15
p − − p̂ −
10 p + − p̂ +
p− − p0
5
p+ − p0
p [bar]

−5

−10

−15
−20 −10 0 10 20 30
θ [CAD]

Figure 11.4 The pressure differences p + − p 0 (blue, solid) and p − − p 0 (red, so-
lid), together with the prediction errors p + − p̂ + (blue, dashed) and p − − p̂ − (red,
dashed) for the pressure curves in Fig. 11.3. Interestingly, the pressure deviation
closely resembled a heat-release rate.

Operating Point 1

20

10
p [bar]

−10

−20

−10 −5 0 5 10 15 20 25
θ [CAD]

Figure 11.5 Pressure-prediction performance at operating point 1. The solid

blue (red) pressure curves correspond to the measured cycle-averaged pressure
change p + − p 0 (p − − p 0 ) due to a positive (negative) θSOI shift of 1 CAD. The das-
hed blue (red) lines correspond to the pressure prediction error for the same θSOI
change.

211
Chapter 11. Pressure Prediction and Efficiency Optimization

Operating Point 2
15

10

5
p [bar]

−5

−10

−15
−10 −5 0 5 10 15
θ [CAD]

Figure 11.6 Pressure-prediction performance at operating point 2. See the figure

caption of Fig. 11.5 for a more detailed description.

Operating Point 3
20

10
p [bar]

−10

−20
−4 −2 0 2 4 6 8 10 12
θ [CAD]

Figure 11.7 Pressure-prediction performance at operating point 3. See the figure

caption of Fig. 11.5 for a more detailed description.

and more steeply for the low-load operating points.

It was here assumed that ∂∆θ/∂θSOI = 1, but when θSOI was decreased
(≈ −25), τ increased, which gave ∂∆θ/∂θSOI < 1. This resulted in an overesti-
mated predicted p change. For late θSOI (≈ −5), τ also increased which gave
∂∆θ/∂θSOI > 1, and resulted in an underestimated p change. This explains the
trends in Figs. 11.5-11.7. It also explains why the R 2 curve was higher for the high

212
 11.3 Results

Prediction Performance
1

0.9

0.8
R 2 [-]

0.7

0.6

0.5
OP 1 OP 2 OP 3
0.4
−4 −2 0 2 4 6 8 10 12 14 16
θ 50 [CAD]

Figure 11.8 R 2 score as a function of θ50 for the three sweeps in Figs. 11.5-11.7.
The pressure-prediction method worked satisfactorily for θ50 ∈ [0, 6], where R 2 ≥
0.9. The performance then started to degrade outside this interval, and more ste-
eply at the low-load operating points.

load experiment in Fig. 11.8, where τ did not change significantly with an almost
constant ∂∆θ/∂θSOI , see Fig. 11.2.

Closed-loop Experiments
This section demonstrates closed-loop performance. Tuning for best perfor-
mance was not carried out as the experiments were focused on convergence and
parameter sensitivity.
Convergence and β - Sensitivity The controller in Algorithm 6 was evaluated
at the investigated operating points. The parameter β and the initial injection ti-
0
ming θSOI were varied to investigate controller convergence. Convergence results
0
are presented in Figs. 11.9-11.11, where θSOI = {20, 10, 0} [CAD] and β = {0.05, 0.2}.
∗
The controller consistently converged to the same θ50 , independently of the
starting point in Figs. 11.9-11.11. The parameter β clearly influenced the conver-
gence rate, where a larger β gave slower convergence. In Fig. 11.10, the controller
behavior in stationarity became oscillatory for β = 0.05. In the same figure, it also
0
seems as if the controller converged faster for β = 0.2 and θ50 = −2. This was cau-
sed by an unintended active p max constraint. The constraint did not affect the
stationary behavior since the estimated d p IMEPg /d ∆θ was zero in the conver-
gence point.
Figures 11.9-11.11 show that 1/β can be viewed as a controller gain, where
the choice of β is a trade-off between convergence speed and stationary
cycle-to-cycle variation. If β was chosen too small, the derivative d p IMEPg /d ∆θ

213
Chapter 11. Pressure Prediction and Efficiency Optimization

Controller Convergence

β = 0.2
β = 0.05
10
θ 50 [CAD]

0 20 40 60 80 100 120 140 160 180

cycle [-]

Figure 11.9 Controller convergence at operating point 1 where the solid lines
correspond to β = 0.05 and the dashed lines to β = 0.2. The controller converged
∗ = 5.8, θ ∗ = −7, in 50 and 150 cycles, depending on β.
to θ50 SOI

Controller Convergence

15 β = 0.2
β = 0.05

10
θ 50 [CAD]

0 20 40 60 80 100 120 140 160 180

cycle [-]

Figure 11.10 Controller convergence at operating point 2, where the solid lines
correspond to β = 0.05 and the dashed lines to β = 0.2. Here, the point of conver-
∗ = 6.4, θ ∗ = −9.5. In the lower dashed θ trajectory, the fast con-
gence is θ50 SOI 50
vergence was due to an unintended active p max constraint. This did however not
affect the point of convergence, since the computed ∂p IMEPg /∂∆θ was zero in the
convergence point.

214
 11.3 Results

Controller Convergence

15 β = 0.2
β = 0.05

10
θ 50 [CAD]

0 20 40 60 80 100 120 140 160 180

cycle [-]

Figure 11.11 Controller convergence at operating point 3, where the solid lines
correspond to β = 0.05 and the dashed lines to β = 0.2. Here, the point of con-
∗ = 5.6, θ ∗ = −9.6, where the controller converged in 5 cycles for
vergence is θ50 SOI
β = 0.05.

caused large controller steps ∆θSOI which led to an oscillatory controller be-
havior. For best performance, β should be increased with load as d p IMEPg /d ∆θ
increases. The point of convergence in Figs. 11.9-11.11 occurred later than the
experimentally found most efficient points in Fig. 11.1. The reason for this will
be discussed in the following section.

∗
Parameter Sensitivity The point of convergence θ50 depends on the parame-
ter values in (11.2), and especially on the parameters of the heat-transfer model.
∗
In order to investigate θ50 sensitivity, the model parameter α (see (11.5)) and the
TDC offset ∆θTDC were varied. These parameters were previously set to α = 5 and
∆θTDC = 0. Convergence results can be viewed in Figs. 11.12 and 11.13 where α
and ∆θTDC were varied stepwise from 1 to 6, and from 2 to -2 CAD, respectively.
Figure 11.12 shows that the magnitude of α affected the convergence point. With
∗
an increased α, θ50 was delayed and the converse was true for a decreased α.
One could view α as a trade-off parameter that weighs efficiency effects from
heat-transfer and exhaust losses, and in that way determines an efficiency opti-
∗
mal θ50 .
∗
In Fig. 11.13, it can be seen that a ∆θTDC of 2 CAD gave a θ50 offset of ap-
proximately 2 CAD. This indicates that the controller accuracy is limited by the
precision of the measured θ, and its synchronization with p.

215
Chapter 11. Pressure Prediction and Efficiency Optimization

α=6
5
α=5
θ 50 [CAD]
α=4

α=3

0
α=2

α=1
100 200 300 400 500
cycle [-]

Figure 11.12 Heat-transfer sensitivity at operating point 1. Here the scaling fac-
tor α was varied stepwise from 1 to 6, which changed the point of convergence.
An α between 3 and 4 would have maximized efficiency according to the data in
Fig. 11.1. In Figs. 11.9-11.11, α = 5 was used.

10 ∆θTDC = 2
θ 50 [CAD]

8 ∆θTDC = 0

6 ∆θTDC = -2

50 100 150 200 250

cycle [-]

Figure 11.13 TDC-offset sensitivity at operating point 1. Here, ∆θTDC was chan-
ged stepwise from 2 to -2 CAD, which indicates that the controller accuracy is li-
mited by the precision of the measured θ and its synchronization with p.

Pressure-Constraint Handling
Controller performance with respect to constraint fulfillment was also investiga-
c c
ted by varying the constraint limits p max and d p max . Result are presented in Figs.
c c
11.14 and 11.15, where p max was varied between 80 to 120 bar and d p max was va-
ried between 20 to 40 bar/CAD. The controller managed to fulfill the constraints
by initially taking a larger positive step in θSOI and then slowly advancing θ50 to
reach the constraint level from below.

216
 11.4 Discussion

120

p max [bar]
100

80

100 200 300 400 500 600 700 800 900

cycle [-]

15
θ 50 [CAD]

10

5
0 100 200 300 400 500 600 700 800 900 1,000
cycle [-]

Figure 11.14 In the upper part of the figure, p max (solid) is displayed together
c
with p max (dashed), The corresponding θ50 is displayed in the lower subdiagram.

With active constraints, the controller was found to enhance cycle-to-cycle

variation. It would therefore be wise to increase β for decreased cycle-to-cycle
variation, if active constraints are expected. The controller also had a specific be-
havior when a constraint was violated, as seen in Fig. 11.14. At cycle 340, the con-
troller delayed θSOI greatly and then slowly advanced combustion timing until
it reached the allowed p max limit. This was because the controller was forced to
c
delay θSOI due to a decrease in p max . The partial derivative ∂p max /∂θSOI was un-
derestimated due to an increase in τ, which explains why the controller took a
too large step to fulfill the constraint.

11.4 Discussion
A desirable combustion timing is a trade-off between exhaust losses, heat trans-
fer, and constraint fulfillment. This chapter introduced a high-level model-based
combustion-timing controller that finds the efficiency-optimal combustion

217
Chapter 11. Pressure Prediction and Efficiency Optimization

50

d p max [bar/CAD]
40

30

20

10
0 100 200 300 400 500 600 700
cycle [-]

20

15
θ 50 [CAD]

10

0 100 200 300 400 500 600 700

cycle [-]

Figure 11.15 In the upper part of the figure, d p max (solid) is displayed together
c
with d p max (dashed). The corresponding θ50 is displayed in the lower subdia-
gram.

timing online. In this way, the controller solves the problem of deciding a
combustion-timing set point, which is necessary for most combustion-feedback
controllers, see for instance Chapter 6 and [Bengtsson et al., 2004; Chiang and
Stefanopoulou, 2005; Widd et al., 2008]. Furthermore, the model-based approach
presented here allows for faster convergence (5 cycles) than data-based extre-
mum seeking controllers that search for efficiency-optimal θ50 using the com-
puted p IMEPg , see [Killingsworth et al. 2009; Lewander et al. 2012]. The method
presented in this chapter could speed up convergence of such methods by provi-
ding an initial guess.
Feedback was introduced through the estimated heat-release rate, which was
utilized by the controller, together with a linearized 0D model, to predict how
the cylinder pressure varies with injection timing. This is a computationally ef-
ficient alternative to that of modeling the relation between fuel injection and
heat-release rate.
The point of convergence was found to be sensitive to TDC offset and

218
 11.5 Conclusions

heat-transfer parameters. The controller should therefore be combined with me-

thods capable of computing these parameters. Examples of such methods were
introduced in Chapter 4. In addition to model parameters, the controller had one
tuning parameter, β. The choice of β was found to be a trade-off between con-
vergence speed and steady-state cycle-to-cycle variation.
It was also found that the pressure prediction performance was dependent on
the assumption of ∂∆θ/∂θSOI as a function of θSOI . In this chapter, ∂∆θ/∂θSOI was
set constant equal to 1. The results indicate that controller performance would
improve if ∂∆θ/∂θSOI were known. An adaptive method could therefore be used
to estimate ∂∆θ/∂θSOI online with θ50 data. A model-based alternative would be
to use a τ model.
The prediction method studied in this chapter will be used in the following
chapter to solve the optimization-problem studied in Chapter 10 with MPC. The
method will also be used to track cycle-resolved pressure set-point trajectories.

11.5 Conclusions
A model-based combustion-timing controller was introduced. The controller
utilized the estimated heat-release rate to predict how the in-cylinder pressure
varies with injection timing. The controller converged close to experimentally
found most efficient combustion timing, and was capable of fulfilling constraints
with respect to p max and d p max .

219
12
Predictive Constraint
Handling and Pressure
Tracking

This chapter investigates how the pressure-prediction method described in the

previous chapter can be used to efficiently fulfill constraints with multiple injec-
tions. This optimal-control problem, defined by (10.10), was previously studied
and solved with a hybrid PI controller in Chapter 10. The approach taken in this
chapter is to instead solve (10.10) on-line with the use of a pressure model and
model predictive control (MPC). An advantage with this approach is that physical
system knowledge is explicitly included in the controller design. The MPC repre-
sentation also has the advantage of being able to handle multiple constraints and
objectives simultaneously, which reduces the controller-design complexity.
A pressure-tracking problem is also considered where the objective is to
follow a cycle-resolved pressure-reference trajectory. The pressure-tracking
problem was previously studied and solved in simulation with iterative lear-
ning control and CFD-based feedforward control for the same purpose [Jörg et
al., 2015; Zweigel et al., 2015]. In contrast to these works, this chapter presents a
model-based pressure tracking controller that does not require a fuel-injection
model.
These two controllers work by the MPC principle, which means that fuel in-
jections are repeatedly optimized with respect to a receding horizon of future
engine cycles. The computations done by the controllers every cycle involves for-
mulating and then solving a quadratic program (QP). The QPs are obtained by ap-
proximating the original optimization problems to obtain real-time compatible
solution times. A heat-release method that separates the estimated heat-release
rate in order to distinguish contributions from different injections was also used,
together with model-based criteria for deciding when to add and remove injec-
tions.

220
 12.1 Model Predictive Control Formulation

The chapter is outlined as follows: The optimal-control problems are presen-

ted and approximated by QPs in Sec. 12.1. In Sec. 12.2, heat-release separation is
used, together with a linearized pressure model to compute expressions needed
to solve the QPs. Additional considerations for changing the number of injections
and avoiding stochastic constraint violation are covered in Secs. 12.3 and 12.4.
Experimental results are presented in Sec. 12.5, and discussion and conclusions
are given in Secs. 12.6 and 12.7.
Data presented in Secs. 12.1 to 12.4 were generated from simulation for illus-
tration purposes whilst the data presented in Fig. 12.5 and Sec. 12.5 were obtai-
ned from engine experiments.

12.1 Model Predictive Control Formulation

The following two optimal-control problems are considered:

1. To efficiently track load and combustion-timing set points subject to con-

straints on peak pressure, pressure-rise rate, NOx and exhaust tempera-
ture.

2. To track a predefined cycle-resolved pressure-reference trajectory.

In both problems, the optimization variable is defined as

³ ´T
1
u = θSOI m 1f ... M
θSOI mM
f (12.1)

where m f denotes the fuel mass, θSOI denotes injection timing, and M is the
number of injections. Injector-current pulse durations θDOI are computed from
m f and the common-rail pressure p rail using an injector map

i
θDOI = M inj (m if , p rail ) (12.2)

The first optimization problem was introduced and motivated in Chap-

ter 10. In the second problem, the controller aims to obtain a predefined
pressure-reference trajectory. This is an alternative control-problem formulation
to combustion-timing and work-output regulation, which is more commonly
used when utilizing pressure-sensor feedback. In both cases, the optimization
problems were reformulated as MPC problems. This procedure will now be pre-
sented in the following sections.

Constraint Fulfillment
The first MPC problem concerns optimization problem (10.10), which was stu-
died in Chapter 10. Here, (10.10) is reformulated as an MPC problem with a pre-
diction horizon and additional costs on control action and load-tracking error.

221
Chapter 12. Predictive Constraint Handling and Pressure Tracking

A constraint on pressure-rise rate is also included

Hp
X
minimize J m f (k) + J p IMEP (k) + J ∆u (k) (12.3)
u(1), . . . , u(H p ) k=1
c
subject to p(θ j , k) ≤ p max θ j = θ0 , . . . , θ f , k = 1, . . . , H p
c
d p(θ j , k)/d θ ≤ d p max θ j = θ0 , . . . , θ f , k = 1, . . . , H p
c
NOx (k) ≤ NOx k = 1, . . . , H p
c T
Tex (k) ≥ Tex k = n Tex , . . . , H p ex
u(k) ∈ U k = 1, . . . , H p

The objective of (12.3) is to minimize a sum of cost functions that penalize

fuel-consumption, load-tracking error and control action: J m f , J p IMEP , and J ∆u ,
with respect to u, over a horizon of H p future engine cycles. The control-action
cost is introduced to obtain robustness by suppressing large changes in u. The
index k denotes engine cycle where k = 0 is the previous engine cycle, H p and
T
H p ex are prediction-horizon lengths in engine cycles.
Constraint limits x c should be fulfilled with respect to pressure p and
pressure-rise rate d p/d θ at sampled crank angles θ j , as well as cylinder-out
NOx emissions and exhaust temperature Tex . Upper bounds on p, d p/d θ and
NOx were motivated by mechanical engine tolerances, engine noise and legi-
slated emission limits. The lower limit for Tex was introduced as a guarantee for
after-treatment system performance. The feasible input set U is described below.
Moreover, the argument u is omitted for the different output variables for ease of
notation.
Cost Functions The fuel-consumption penalty J m f (k) in (12.3) is represented
r,i
as a cost on combustion-timing deviation from efficient set points θCT , for the
different injections i = 1, . . . , M
M
X r,i
J m f (k) = α m if (k)(θCT i
− θCT (k))2 (12.4)
i =1

which allows for control of a multimodal heat-release rate. An efficiency-optimal

combustion timing is located somewhere after TDC, and it is here assumed that
r,i
θCT is obtained from engine experiments or computed using the method presen-
ted in Chapter 11. The penalty terms in (12.4) are weighted with m if to prioritize
efficient combustion timings for heavier injections. The fuel-mass weights also
introduce the controller behavior of moving fuel from a less efficient injection to
a more efficient injection.
r
The engine load p IMEP should follow a set point p IMEP . This objective is represen-
ted by the quadratic cost
r
J p IMEP (k) = β(p IMEP − p IMEP (k))2 (12.5)

222
 12.1 Model Predictive Control Formulation

where p IMEP is used to denote p IMEPg in this chapter. The control-action cost
term in (12.3) is given by

J ∆u (k) = (u(k) − u(k − 1))T R(u(k) − u(k − 1)) (12.6)

where u(0) is the previous-cycle injection configuration.

Pressure Tracking
The MPC for pressure tracking is given by

Hp
X
minimize J p (k) + J ∆u (k) (12.7)
u(1), . . . , u(H p ) k=1

subject to u(k) ∈ U, k = 1, . . . , H p

where the cost J p penalizes in-cylinder pressure deviation from a predefined

pressure-reference trajectory.
Cost Functions The first cost term in (12.7) is given by

θf
X ¡ ¢2
J p (k) = αt r p r (θ j ) − p(θ j , k) (12.8)
θ j =θ0

which is the sum of squared p deviations from p r at sampled crank angles θ j .

The summation interval was set to only penalize pressure deviation close to TDC
and during the expansion stroke, since this is the part which is directly controlled
by fuel injection. The control-action cost term in (12.7) is given by (12.6).
Pressure Reference The pressure reference p r in (12.8) is parameterized as an
ideal limited-pressure cycle, where a portion of the fuel αv is burned with con-
r
stant cylinder volume at the start of combustion θSOC . The remainder of the fuel
αp is then burned at constant cylinder pressure. Examples of p r are presented
r
in Fig. 12.1 for different αv . The start of combustion θSOC , αv , and the desired
r
total fuel-energy burned Q c are considered to be tuning parameters that de-
termine p r . A more detailed description of ideal engine cycles can be found in
[Heywood, 1988].

Input Constraints
Input constraints were enforced through saturation in Chapter 10. Here, the in-
put constraints are instead included in the optimization problem, where con-
straints on u are imposed in order for the solution to be realizable. Injected fuel
masses have to be positive

m if ≥ 0, i = 1, . . . , M (12.9)

223
Chapter 12. Predictive Constraint Handling and Pressure Tracking

140
αv = 1
αv = 0.5
120
αv = 0

100

80
p r [bar]

60

40

20

0
−20 −10 0 10 20 30 40 50
θ [CAD]

Figure 12.1 The pressure reference signal p r is an ideal limited-pressure cycle,

where a portion of the fuel αv is burned at constant cylinder volume at the start of
combustion θSOCr , and the remainder of the fuel αp is burned at constant cylinder
r
pressure. The constant-volume fraction αv is here varied from 0 to 1, with θSOC =
0 and Q cr = 5000 J.

Furthermore, the fuel-injection pulses have to be positioned so that they do not

overlap. This imposes the following constraint
i −1 i −1 i
θSOI + θDOI + θh ≤ θSOI , i = 2, . . . , M (12.10)
where θDOI is computed with (12.2) and θh is a margin related to the hydraulic
delay of the injector, allowing the injector to close in-between injections. Addi-
tional absolute and relative constraints on u
u l ≤ u(k) ≤ u u , k = 1, . . . , H p
(12.11)
∆u l ≤ u(k) − u(k − 1) ≤ ∆u u , k = 1, . . . , H p
were added to provide robustness and limit the controller operating range. The
set of injection configurations fulfilling (12.9)-(12.11) are denoted U in (12.3) and
(12.7).

Cost-Function Weights
The cost-function weights α, αt r , β and the diagonal matrix
³ ´
1 1 M M
R = diag R m f
R SOI . . . Rm f
R SOI >0 (12.12)

224
 12.1 Model Predictive Control Formulation

are parameters that determine the trade-off between control action and fulfill-
ment of the different tracking objectives. Suitable values will be presented toge-
ther with experimental results in Sec. 12.5.

Problem Approximation
We have now obtained two nonlinear-programming problems on the form

minimize J (U ) (12.13)
U
subject to f (U ) ≤ c
U ∈U

where
¡ ¢T
U = u(1) ... u(H p ) (12.14)

The two functions f (U ) and J (U ) are non-linear in U , and f demands extensive

modeling. Numerical methods could of course be used to search for local optima
of (12.13). However, this would require multiple evaluations of J and f which
is computationally expensive, since it involves solving differential equations. To
allow for shorter computational times, necessary for on-line applications, it was
decided to approximate (12.13) by a QP, at the current input u 0

1
minimize J (u 0 ) + ∇J (u 0 )T ∆U + ∆U T ∇2 J (u 0 )∆U (12.15)
∆U 2
subject to f (u 0 ) + ∇ f (u 0 )T ∆U ≤ c
u 0 + ∆U ∈ Û

where ∆U is the deviation from u 0

¡ ¢T ¡ ¢T
∆U = u(1) ... u(H p ) − u 0 = ∆u(1) ... ∆u(H p ) (12.16)

Another motivation for approximating (12.13) was that gradients and Hessians
can be computed using simpler physical models, valid for smaller variations in u.
The optimization problem can then be re-approximated on a cycle-to-cycle ba-
sis when new measurements have been obtained. This means that the gradients
∇ f (u 0 ), ∇J (u 0 ) and Hessian ∇2 J (u 0 ) have to be recomputed as u 0 changes. The
following section presents how to obtain gradients and Hessians with the use of
physical models. A linear approximation of Û was obtained by linearizing M inj

i −1 i −1 i −1
∂M inj (u 0 )
θSOI (0) + ∆θSOI + θDOI (0) + ∆m if −1 + θh ≤ θSOI
i i
(0) + ∆θSOI (12.17)
∂m f

225
Chapter 12. Predictive Constraint Handling and Pressure Tracking

12.2 Modeling and Heat-Release Detection

The proposed method for computing the gradients and Hessian in (12.15) is
to first establish a relationship between ∆u and variations in dQ c /d θ. For this
purpose, the estimated dQ c /d θ at the previous engine cycle k = 0 is utilized.
The resulting effects of ∆u on p, NOx and Tex are then computed using the
pressure-linearization method presented in Chapter 11.

From ∆u to d ∆Q c /d θ
First, the previous cycle dQ c /d θ is obtained from

dQ c γ dV 1 d p dQ ht
= p + V + (12.18)
dθ γ − 1 dθ γ − 1 dθ dθ

If multiple injections are used, we have to separate dQ c /d θ among the injec-

tions. It is assumed that dQ c /d θ consists of the heat-release generated from the
different injections according to

dQ c X M dQ i
c
= (12.19)
dθ i =1 d θ

The procedure for obtaining dQ ci /d θ from dQ c /d θ with M injections was pre-

sented in Chapter 4. It is now briefly recaptured for clarity. First, the M most signi-
ficant peaks are located, where detected peaks have to be larger than a threshold
dQ t and separated with a minimum distance θd . The detected peaks constitute
i
the combustion timings θCT in (12.4).
With the peaks detected, dQ c /d θ can be separated in different intervals

 dQ c
d Q̂ c 
i if l i ≤ θ ≤ d i
= dθ (12.20)
dθ 
0 otherwise

where the bounds l i and d i are determined by the minima between the detec-
ted peaks. The obtained heat-release rates d Q̂ ci /d θ are then smoothed with a
zero-phase filter to obtain more physical heat-release shapes. Finally, d Q̂ ci /d θ
have to be normalized so that (12.19) is fulfilled

dQ ci ³XM d Q̂ i ´−1 d Q̂ i dQ
c c c
= (12.21)
dθ i =1 d θ d θ d θ

The four steps of the detection procedure are illustrated in Fig. 12.2. In the case
of detecting less than M peaks, which is possible for small injections or if com-
bustion from different injections overlap, the detected peaks are allocated to the
largest injections. Furthermore, the detection procedure accounts for the orde-
ring of injections when allocating combustion timings to injections.

226
 12.2 Modeling and Heat-Release Detection

600 600
dQ c /d θ [J/CAD] 1. 2.

dQ c /d θ [J/CAD]
400 400

200 200

0 0
−20 0 20 −20 0 20
θ [CAD] θ [CAD]

600 600

3. 4.
dQ c /d θ [J/CAD]

dQ c /d θ [J/CAD]
400 400

200 200

0 0
−20 0 20 −20 0 20
θ [CAD] θ [CAD]

Figure 12.2 The method used for separating dQ c /d θ among the different in-
jections (from 1 to 4). First, the heat-release rate is obtained from the measured
pressure signal (1) and the most significant M peaks are detected (2). dQ c /d θ
is then separated in different intervals according to the peak locations (3). The
heat-release rates d Q̂ ci /d θ in the different intervals are then filtered and normali-
zed (4).

We can now relate a change in u, ∆u, to a change in dQ c /d θ, d ∆Q c /d θ. A change

i
in θSOI is assumed to give a crank-angle shift in the part of dQ c /d θ that is affected
by injection i , dQ ci /d θ. This gives the partial derivatives

∂ dQ c d 2Q ci
=−
i
∂θSOI dθ d θ2
(12.22)
i
∂θCT
i
=1
∂θSOI

A change in m if is assumed to only affect the accumulated heat-released in

227
Chapter 12. Predictive Constraint Handling and Pressure Tracking

600 600
dQ c /d θ [J/CAD]

dQ c /d θ [J/CAD]
400 400

200 200

0 0
−20 0 20 40 −20 0 20 40
θ [CAD] θ [CAD]

Figure 12.3 The assumed relation between changes in m if (left) and θSOI
i (right)
to changes in dQ c /d θ. An increase in m if results in an increase in accumulated
dQ ci /d θ, and a shift in θSOI
i results in a shift in dQ ci /d θ.

dQ ci /d θ, which gives
∂ dQ c Q LHV dQ ci
= R (12.23)
∂m if d θ dQ ci d θ

These assumptions are illustrated in Fig. 12.3. In both cases, the shape of dQ c /d θ
is preserved. It is also assumed that ∆u j and dQ ci /d θ are decoupled when i 6= j
and that ignition delays remain constant. These are approximations since subse-
quent injections are coupled both through ignition delay and rail pressure. The
shape of dQ c /d θ is also known to change slightly with m f and θSOI . If more accu-
rate combustion models are available, those could be incorporated for potential
controller-performance improvement. The approximations made could still be
motivated for small changes in ∆u, since feedback from subsequent engine cy-
cles corrects for unmodeled effects. Second derivatives, necessary for computing
the Hessians in (12.15) are given by

∂2 dQ c d 3Q ci
=
i
∂(θSOI )2 dθ d θ3

∂2 dQ c Q LHV d 2Q ci
= − R (12.24)
i
∂θSOI ∂m if d θ dQ ci d θ 2

∂2 dQ c
=0
∂(∂m if )2 dθ

228
 12.2 Modeling and Heat-Release Detection

To summarize, we have derived the following quantities relating ∆u to ∆dQ c /d θ

µ ¶T
dQ c Q LHV dQ c1 d 2Q c1 Q LHV dQ cM d 2Q cM
∇ = R 1 2
... R
dθ dQ c d θ dθ dQ cM d θ d θ2
 

 Q LHV d 2Q ci 
 (12.25)

 0 − R 
dQ  dQ i d θ2 



2 c c
∇ = diag  
dθ 
  Q LHV d 2Q ci d 3Q ci



 

 −R i d θ2 3 
dQ c dθ

where ∇2 dQ c /d θ is a block-diagonal matrix.

From d ∆Q c /d θ to ∆p
The relation between d ∆Q c /d θ and ∆p can now be established with the lineari-
zed pressure model introduced in Chapter 11
µ ¶
d ∆p γ dV d µ(p 0 , θ) γ − 1 d ∆Q c
=− + ∆p + (12.26)
dθ V dθ dp V dθ

where ∆p is the first-order deviation from the initial pressure p 0 due to d ∆Q c /d θ,

which is related to ∆u through

µ ¶
d ∆Q c dQ c T
= ∇ ∆u (12.27)
dθ dθ

The pressure deviation ∆p is given by the solution to (12.26)

Z θ d ∆Q c (ϑ)
∆p(θ) = Φ(θ, ϑ)Γ(ϑ) dϑ (12.28)
θIVC dϑ

where
µ Z θ ¶µ ¶γ
d µ(p 0 , τ) V (ϑ)
Φ(θ, ϑ) = exp − dτ
ϑ dp V (θ)
(12.29)
γ−1
Γ(ϑ) =
V (ϑ)

Figure 12.4 shows how ∆p is related to ∆u (dashed), together with p 0 (solid,

black), and p at u 0 + ∆u (solid, blue), obtained by solving the nonlinear model
(11.2). Note that there is a deviation between the linear approximation and the
solution to (11.2).

229
Chapter 12. Predictive Constraint Handling and Pressure Tracking

100
p0
p 0 + ∆p
80 p(u 0 + ∆u)

60
p [bar]

40

20

0
−20 −10 0 10 20 30 40 50
θ [CAD]

Figure 12.4 An illustration of how ∆u affects ∆p through (12.28) (dashed). The

initial pressure p 0 (solid, black) is also presented together with p at u 0 +∆u, com-
puted with the nonlinear model (11.2) (solid, blue).

With (12.28), the gradients with respect to p, p IMEP and d p/d θ are given by
Z θ µ ¶
dQ c (ϑ)
∇p = Φ(θ, ϑ)Γ(ϑ) ∇ dϑ
θIVC dϑ
Z θEVO
1 (12.30)
∇p IMEP = ∇pdV
Vd θIVC

d p d ∇p
∇ =
dθ dθ
Hessians are given by
Z θ µ ¶
2 2 dQ c (ϑ)
∇ p= Φ(θ, ϑ)Γ(ϑ) ∇ dϑ
θIVC dϑ
(12.31)
Z θEVO
2 1 2
∇ p IMEP = ∇ pdV
Vd θIVC

Note that the gradients and Hessians above, except for ∇p IMEP and ∇2 p IMEP ,
are functions of θ.

NOx
The NOx constraint was linearized using the NOx -formation model in Sec 2.5.
Since this model is not easily linearized, partial derivatives of the cylinder-out

230
 12.2 Modeling and Heat-Release Detection

10
Data Model
5
Tex [◦ C]
0

−5

−10
0 200 400 600 800 1,000 1,200 1,400 1,600 1,800
cycle [-]

200
TEVO [◦ C]

−200

0 200 400 600 800 1,000 1,200 1,400 1,600 1,800

cycle [-]

Figure 12.5 Experimental Tex and TEVO data (solid) together with (12.33) output
(dashed) when varying θSOI . The temperature scales are relative to steady-state
values at p IMEP = 5 bar.

NOx concentration with respect to u was obtained by solving (2.73) and applying
numerical differentiation

∂NOx NOx (u 0 + ∆m if ) − NOx (u 0 )

≈
∂m if ∆m if
(12.32)
i
∂NOx NOx (u 0 + ∆θSOI ) − NOx (u 0 )
i
≈ i
∂θSOI ∆θSOI

In the forward step, NOx (u 0 + ∆u) was computed by solving (2.73) with modified
cylinder pressures, temperatures and heat-release rates p(u 0 + ∆u), T (u 0 + ∆u),
dQ c /d θ(u 0 + ∆u), obtained from the linearized expressions presented above.
With M injections, this amounts to solving (2.73) 2M + 1 times. The most
computationally-demanding part of computing NOx was to compute gas pro-
perties as a function of temperature. To reduce the computational load, it was
decided to also use the gas properties computed at u 0 in the forward steps.

231
Chapter 12. Predictive Constraint Handling and Pressure Tracking

Exhaust Temperature
Examples of lumped-parameter Tex models, suitable for control applications are
presented in [Eriksson and Nielsen, 2014]. This methodology was adopted here
and used to model the relation between the temperature at exhaust-valve ope-
ning (EVO) TEVO and Tex

∆Tex (k + 1) = ΦTex ∆Tex (k) + ΓTex ∆TEVO (k) (12.33)

where ∆Tex (k) and ∆TEVO (k) are temperature deviations from an equilibrium
0 0
point Tex , TEVO , and ΦTex and ΓTex are model parameters. Differentiation with
respect to TEVO
V (θEVO )TIVC
∇TEVO = ∇p(θEVO ) (12.34)
p IVC VIVC
establishes a relation with ∇p(θEVO ), which is given by (12.30).
Experimental Tex and TEVO data (solid) during θSOI step changes are presen-
ted together with (12.33) output (dashed) in Fig. 12.5. The temperature scales are
relative to steady-state values at p IMEP = 5 bar. Values for ΦTex and ΓTex were ob-
tained from engine data and the MATLAB system-identification toolbox.
In order to incorporate the long time constants of Tex into the MPC pro-
blem formulation, exhaust temperature was predicted over a longer horizon
T
H p ex = 500 and with a longer sampling interval, (n Tex = 100 cycles). Moreover, the
model was augmented with a disturbance state d TEVO to keep track of the model
steady-state offset

∆Tex (k + 1) = ΦTex ∆Tex (k) + ΓTex (∆TEVO (k) + d TEVO (k))

(12.35)
d TEVO (k + 1) = d TEVO (k)

The Kalman filter presented in Chapter 3 was then used to estimate d TEVO using
(12.35) and Tex measurements.

QP Approximations
Constraint Fulfillment With gradients and Hessians available, the QP approxi-
mation of (12.3) is given by
Hp
X
minimize ˆ ˆ
J θCT (k) + J p IMEP (k) + J ∆u (k) (12.36)
∆U k=1
0 c
subject to p max + ∇p 0 (θ j )∆u(k) ≤ p max θ j = θ0 , . . . , θ f , k = 1, . . . , H p
0 c
d p max + ∇d p 0 (θ j )/d θ∆u(k) ≤ d p max θ j = θ0 , . . . , θ f , k = 1, . . . , H p
NO0x + ∇NO0x ∆u(k) ≤ NOcx k = 1, . . . , H p
0 c T
Tex + ∆Tex (k) ≥ Tex k = n Tex , . . . , H p ex
u 0 + ∆u(k) ∈ Û k = 1, . . . , H p

232
 12.3 Adding and Removing Injections

The cost function JˆθCT is given by

 1 T  1 
fθ Hθ 0 0
 .CT   CT

 
JˆθCT (k) = α  ..  ∆u(k) + ∆u T (k) 
 0
..
.

0  ∆u(k) (12.37)
 
fM 0 0 HθM
θCT CT

where  
r,i i ,0 2
(θCT − θCT )
f θiCT =  
−2m if ,0 (θCT
r,i i ,0
− θCT )
  (12.38)
r,i i ,0
0 −(θCT − θCT )
Hθi CT = α  r,i i ,0

−(θCT − θCT ) m if ,0
Furthermore, the cost function Jˆp IMEP (k) is given by
r
¡ T
¢
Jˆp IMEP (k) = −β(p IMEP − p IMEP,0 ) 2∇p IMEP,0 ∆u(k) + ∆u T (k)∇2 p IMEP,0 ∆u(k)
+ β∆u T (k)∇p IMEP,0 ∇p IMEP,0
T
∆u(k) (12.39)
Index 0 in (12.36) to (12.39) above denotes cycle k = 0.
Pressure Tracking The QP approximation (12.7)
Hp
X
minimize Jˆp (k) + J ∆u (k) (12.40)
∆U k=1
subject to u 0 + ∆u(k) ∈ U, k = 1, . . . , H p
is obtained by computing the approximated pressure-tracking cost
θf
X ¡ ¢
Jˆp (k) = − αt r (p r (θ j ) − p 0 (θ j )) 2∇p 0T (θ j )∆u(k) + ∆u T (k)∇2 p 0 (θ j )∆u(k)
θ j =θ0

+ αt r u T (k)∇p 0 (θ j )∇p 0T (θ j )∆u(k) (12.41)

With the QP approximations of (12.3) and (12.7) defined, the following two sec-
tions cover additional logic for adding and removing injections and practical mo-
difications for constraint fulfillment.

12.3 Adding and Removing Injections

The optimal number of injections changes depending on operating conditions.
This was shown in Chapter 10 and is also illustrated in Fig. 12.6 where optimal
injection configurations for two different p r are presented. This fact presents the
problem of when to add and remove injections. Two different approaches used
for (12.36) and (12.40) are presented below.

233
Chapter 12. Predictive Constraint Handling and Pressure Tracking

100 100
p [bar]

p [bar]
50 50

0 0
−20 0 20 40 −20 0 20 40
θ [CAD] θ [CAD]

Figure 12.6 The optimal number of injections changes depending on the ope-
rating conditions. To the right, a single fuel injection is optimal for tracking the
constant-volume combustion reference pressure (dashed). To the left, it is instead
optimal with three fuel injections for tracking of a limited-pressure cycle (dashed).

Constraint Fulfillment
The simulation results presented in Chapter 10 suggested the following logic for
adding and removing injections when solving (12.3) with up to three injections,
pilot, main and post:

• Add a pilot injection if any of the predicted outputs NOx , p max or d p/d θ
are larger than x c − ²add
x .
c
• Add a post injection if the predicted Tex is smaller than Tex + ²add
Tex .

• Remove the pilot injection if predicted NOx , p max and d p/d θ are smaller
than x c − ²rem 1
x , and the pilot-fuel mass m f is sufficiently low.

c
• Remove the post injection if predicted Tex is larger than Tex + ²rem
Tex , and the
post-fuel mass m 3f is sufficiently low.

The constraint margins should fulfill ²rem

x > ²add
x > 0 to avoid limit cycles.
Adding and removing injections introduces disturbances in p IMEP and subse-
quent θCT . Compensation on the form
2 2
θSOI (k + 1) = θSOI (k) ± ∆θ adj
adj (12.42)
m 2f (k + 1) = m 2f (k) ± ∆m f

2
was therefore introduced. The compensation in (12.42) adjusts θSOI (k + 1) when
a pilot injection is added or removed to compensate for variation in ignition de-

234
 12.3 Adding and Removing Injections

100 100
p [bar]

p [bar]
50 50

0 0
−20 0 20 40 −20 0 20 40
θ [CAD] θ [CAD]

Figure 12.7 An illustration of how additional inactive injections were modeled

using extrapolation. Here, the extrapolation is performed when p r changes from
a constant-volume cycle (left) to a limited-pressure cycle (right). Two additional
injections are extrapolated from the currently active single-injection heat-release
rate.

lay, and adjusts m 2f (k + 1) when a pilot or post injections is added or removed to

compensate for variation in p IMEP .

Pressure Tracking
A different method was adopted in (12.40) for adding and removing injections.
The method used for adding injections was to extrapolate from already active in-
jections. For example, if injection i is inactive, ∆u with respect to injection i is as-
sumed to correspond to variations of a shifted (∆θ) and rescaled (x i ) heat-release
rate dQ ca /d θ, corresponding to an already active injection

dQ ci dQ ca
= xi (θ − ∆θ) (12.43)
dθ dθ

Extrapolation was conducted when p r changed significantly, as presented in

Fig. 12.7. Inactive injections were then turned on if the solution to (12.40) sug-
gested that m if should be increased. An alternative solution would be to instead
solve (12.40) with and without extrapolated injections separately, and add injec-
tions if the corresponding cost is significantly lower. Injections were turned off
once the controller reached sufficiently low m if .

235
Chapter 12. Predictive Constraint Handling and Pressure Tracking

12.4 Constraint Handling

To guarantee feasible solutions to (12.36), output constraints were softened by
introducing variables ²x

x 0 + ∇x 0T ∆u(k) ≤ x c + ²x
(12.44)
²x ≥ 0

and adding corresponding terms

ρ x ²2x (12.45)
to the cost function to penalize constraint violation. This modification guaran-
tees feasible solutions with respect to output constraints even if the constraints
can not be fulfilled. By choosing ρ x sufficiently large, the solution will however
always fulfill the output constraint if possible [Maciejowski, 2002]. Constraint
margins x σ were also introduced for constraint fulfillment despite stochastic
cycle-to-cycle variation

x 0 + ∇x 0 ∆u(k) ≤ x c + ²x + 2x σ
(12.46)
²x ≥ 0

The constraint margins were here pre-computed from measured output stan-
dard deviations, meaning that the margins provide probabilistic guarantees for
constraint fulfillment. An alternative approach would be to estimate x σ on-line.
The reason for introducing constraint margins was to remediate the problem of
stochastic constraint violation that occured in previous chapters.

12.5 Experimental Results

An experimental controller evaluation was conducted to test the controller per-
formance in transient operation. The evaluation was done by testing the ability
r
of MPC (12.36) to fulfill different constraints during p IMEP step changes. MPC
r
(12.40) was tested by varying p parameters. Experimental operating conditions
are presented in Table 12.1. In some of the experiments, it was decided not to
turn off the pilot injection completely to suppress the pressure-rise rate of the
main-injection.

Constraint Fulfillment
This section presents experimental results for MPC (12.36). The controller was
r r
tuned for θCT - and p IMEP -tracking response times within 10 engine cycles. Con-
troller gains and input constraints used are presented in Table 12.2. Gain sche-
duling was implemented by increasing R SOI (×4) for the controller to vary θSOI
more cautiously in the vicinity of a constraint limit. A prediction horizon of two

236
 12.5 Experimental Results

Table 12.1 Operating Conditions

p IMEP 4-10 bar

Nspeed 1200 rpm
p rail 1000 bar
EGR 0-30 %
fuel 80/20 vol%
gasoline/n-heptane

engine cycles H p = 2 was used. This enabled mean QP-solver computation times
of 2 ms. The QP was constructed below 50 ms, where ≈ 90% of the time was used
to compute the NOx gradients in (12.32). Fulfillment of each constraint was eval-
r
uated separately during p IMEP step changes. Experimental results are presented
in the following sections.

p max Handling of p max constraints was investigated by letting the controller

r c
follow p IMEP step changes from 5 to 10 bar with p max = 80 bar, see Fig. 12.8. This
c
is far from the real p max of the engine but was used here to illustrate the controller
behavior.
r
As p IMEP was increased from 5 to 10 bar with a single injection, θCT = 8 CAD
c
could not be maintained without violating p max = 80 bar. The controller there-
fore acted by increasing m 1f at cycle 290, as p max
c
was approached. This resulted
1 2
in an additional detected θCT shortly after TDC. Moreover, θSOI had to be delayed
r
8 CAD for constraint fulfillment. As p IMEP decreased, the controller returned to a
r
single-injection configuration with θCT = θCT . The reason for removing the pilot
injection when the constraint became inactive was because the controller prio-
r
ritized the larger main injection to follow θCT due to the higher main-injection
cost-function penalty. This controller behavior was a result of weighting θCT er-
rors with m if in (12.4). The pilot injection was then turned off once m 1f was suffi-
ciently low. The dashed line in the mid-left subdiagram in Fig. 12.8 indicates the
level for which the pilot injection was turned off if no output constraints were
active.

Table 12.2 MPC weights and constraints for (12.36), units for θSOI and m f are
[CAD] and [mg].

α = 0.0005 β = 400 R θSOI = 5 Rm f = 2

−2.5 ≤ ∆θSOI ≤ 2.5 −10 ≤ ∆m f ≤ 10 −25 ≤ θSOI ≤ 20 0 ≤ m f ≤ 180

237
Chapter 12. Predictive Constraint Handling and Pressure Tracking

20 1
θSOI 2
θSOI 3
θSOI
20

θ SOI [CAD]
θ CT [CAD] 10
0
0
−20
100 200 300 100 200 300
cycle [-] cycle [-]

150 12

10

p IMEP [bar]
100
m f [mg]

8
50
6

0 4
100 200 300 100 200 300
cycle [-] cycle [-]

90
p max [bar]

80

70

60
50 100 150 200 250 300
cycle [-]

Figure 12.8 Evaluation of p max -constraint fulfillment. The controller was set to
r
follow p IMEP step changes from 5 to 10 bar. As p IMEP was increased from 5 to 10
r = 8 CAD could not be maintained without viola-
bar with a single injection, θCT
c
ting p max = 80 bar. The controller acted by increasing m 1f as the constraint was
1 (red) shortly after TDC.
approached. This resulted in an additional detected θCT
2 (blue) was delayed 8 CAD for constraint fulfillment. Crank-angle
Moreover, θSOI
resolved data for cycles 45 and 80 are presented in Fig. 12.9.

Crank-angle resolved data for cycles 45 (left) and 80 (right) are presented in
Fig. 12.9. In-cylinder pressure, injector current and dQ c /d θ are presented toge-
r
ther with vertical lines indicating θCT and θCT . The dashed pressure curves vi-
sible at cycle 45 are predicted cylinder pressures for two subsequent engine cy-
r
cles. The increase in predicted pressure was due to the increase in p IMEP . Two
1 2
θCT are detected at cycle 80 due to the increase in m f and θSOI . The red and

238
 12.5 Experimental Results

80 80
cycle 45 cycle 80
60 60
p [bar]

p [bar]
40 40

20 20

0 0
−20 0 20 40 −20 0 20 40
θ [CAD] θ [CAD]

Figure 12.9 In-cycle data for cycles 45 (left) and 80 (right) in Fig. 12.8. In-cylinder
pressure, injector current and dQ c /d θ are presented together with the vertical
r . The dashed pressure curves visible at cycle 45 are
lines indicating θCT and θCT
predicted cylinder pressures for two subsequent engine cycles. The increase in
r
predicted pressure was due to the increase in p IMEP . At cycle 80, it can be seen
how the controller separated dQ c /d θ between the pilot and main injection.

blue heat-release rates indicate how the combustion detection method separa-
ted dQ c /d θ into dQ c1 /d θ and dQ c2 /d θ.
d p/d θ The d p/d θ limit was varied at constant p IMEP in the experiment pre-
1,2
sented in Fig. 12.10. The controller increased m 1f and delayed θSOI c
as d p max was
decreased. The pilot and main fuel burned simultaneously and only one com-
bustion timing was therefore detected, meaning that dQ c /d θ was attributed to
and controlled by the main injection. Disturbances in p IMEP are visible when m 1f
changed. These disturbances could have been attenuated by improving the ca-
adj
libration of ∆m f in (12.42). The d p/d θ limit was occasionally violated due to
cycle-to-cycle variation.
In-cycle data for cycles 15 (left) and 150 (right) are presented in Fig. 12.11. The
c
solid and dashed tangents correspond to d p/d θmax and d p max , respectively. The
1 2 c
pilot fuel amount m f was increased and θCT was delayed to fulfill d p max . The
injected fuel burned simultaneously, which resulted in one combustion timing
detected.
r
NOx The controller was once again set to follow p IMEP step changes to evaluate
NOx -constraint handling where the solution to (2.73) was used as a virtual NOx
r 2
sensor, see Fig. 12.12. As p IMEP increased, θCT was delayed and m 1f increased for
NOx to remain below NOcx = 500 ppm. Constraint violation occurred due to NOx
overshoots as p IMEP was increased. The constraint was however later fulfilled in
steady state. The NOx overshoot could possibly be reduced by adding pilot fuel

239
Chapter 12. Predictive Constraint Handling and Pressure Tracking

20 20 1
θSOI 2
θSOI 3
θSOI
10

θ SOI [CAD]
θ CT [CAD] 10
0

0 −10

−20
100 200 300 100 200 300
cycle [-] cycle [-]

150 10

p IMEP [bar]
100
m f [mg]

50
6
0
100 200 300 100 200 300
cycle [-] cycle [-]

30
d p max [CAD/bar]

20

10

0
50 100 150 200 250 300
cycle [-]

c
Figure 12.10 Evaluation of d p/d θ constraint fulfillment where d p max was va-
1 1,2
ried at constant p IMEP . The pilot fuel amount m f was increased and θSOI were
delayed to fulfill the constraint. The pilot and main fuel burned simultaneously
and only one combustion timing was detected, meanin87g that dQ c /d θ was at-
tributed to and controlled by the main injection. Disturbances in p IMEP are visible
when m 1f changes around cycle 60 although compensation in m 2f was made.

more carefully, allowing for larger θSOI changes, and/or having a longer predic-
tion horizon. Small NOx overshoots could however be acceptable if NOcx was set
to regulate accumulated NOx emissions.
In-cycle data for cycles 25 (left) and 65 (right) from the experiment in
Fig. 12.12 are presented in Fig. 12.13. The solid and dashed purple lines corres-
pond to in-cylinder NOx and NOcx , respectively. The pilot-fuel mass m 1f was in-
2
creased and θCT delayed at cycle 65, in order for NOcx = 500 ppm to be fulfilled.

240
 12.5 Experimental Results

80 80
cycle 15 cycle 150
60 60
p [bar]

p [bar]
40 40

20 20

0 0
−20 0 20 40 −20 0 20 40
θ [CAD] θ [CAD]

Figure 12.11 In cycle data for cycles 15 (left) and 150 (right) from the experiment
c
in Fig. 12.10. The solid and dashed tangents correspond to d p/d θmax and d p max
1 2
respectively. The controller increased m f and delayed θCT at cycle 150 to fulfill
c
d p max = 10 bar/CAD.

c
T ex Figure 12.14 shows how the controller managed to keep Tex above Tex =
◦ r
240 C , whilst p IMEP was varied between 5 and 7 bar. The controller introduced a
c
post injection as Tex approached Tex . The post-injection mass m 3f was then used
c
to regulate Tex above Tex . A p IMEP disturbance is visible when the post injection
was introduced at cycles 180 and 750.
In-cycle data from cycles 5 (left) and 350 (right) in Fig. 12.14 are presented
c
in Fig. 12.15. A post injection was introduced at cycle 350 to fulfill Tex = 240◦C
2 r 3
whilst θCT was kept at θCT . The post-injection combustion timing θCT and cor-
responding heat-release rate dQ c3 /d θ (green) were detected by the controller.

Summary
It can be concluded that the controller was able to fulfill the different constraints
as intended. Speed of convergence was higher than for the heuristic control-
ler design in Chapter 10 with 10-20 cycles as compared to 40-50 cycles. The
model-based controller presented here was also better at avoiding constraint
violation. Both because of its predictive capability, and because of its constraint
margins.

Pressure Tracking
This section presents experimental results with MPC (12.40). The controller was
investigated by varying the pressure-reference parameters αv , θSOC and Q cr . Con-
troller parameters used are presented in Table. 12.3.
The ability to follow αv changes with two injections is presented during a
10-cycle transition from p 1r with αv = 0.5 (red) to p 2r with αv = 0.2 (blue) in

241
Chapter 12. Predictive Constraint Handling and Pressure Tracking

20 20 1
θSOI 2
θSOI 3
θSOI
10

θ SOI [CAD]
θ CT [CAD] 10
0

0 −10

−20
200 400 200 400
cycle [-] cycle [-]

150 10

p IMEP [bar]
100 8
m f [mg]

50 6

0 4
200 400 200 400
cycle [-] cycle [-]

800

600
NOx [ppm]

400

200

100 200 300 400 500

cycle [-]

r
Figure 12.12 Evaluation of NOx -constraint fulfillment. As p IMEP 2
increased, θCT
was delayed and m 1f increased for the constraint to be fulfilled. Constraint viola-
tion occured as p IMEP increased. The constraint was later fulfilled in steady state.

Fig. 12.16. With αv = 0.5, the controller had a relatively large m 1f . With more
constant-pressure combustion, the controller increased m 2f and delayed θSOI
1

2
and θSOI .
Figure 12.17 presents in-cycle data during a transition from p 1r with Q cr = 3000
J (red) to p 2r with Q cr = 6000 J (blue) with one injection. When Q cr was increased,
the controller increased m f whilst θCT was kept constant.
r
A θSOC transition from p 1r with θSOC
r
= 10 CAD (red) to p 2r with θSOC
r
= 5 CAD
r
(blue) is presented in Fig. 12.18. When θSOC was advanced, the controller adjus-
r
ted θSOI whilst m f was kept constant.

242
 12.5 Experimental Results

80 80
cycle 25 cycle 65
60 60
p [bar]

p [bar]
40 40

20 20

0 0
0 20 40 60 0 20 40 60
θ [CAD] θ [CAD]

Figure 12.13 In-cycle data for cycles 25 (left) and 65 (right) from the experiment
in Fig. 12.12. The solid and dashed purple lines correspond to in-cylinder NOx
and NOcx respectively. The pilot-fuel mass m 1f was increased and θCT 2 delayed at

cycle 65 for NOcx = 500 ppm to be fulfilled.

Data illustrating the suggested method for introducing injections during pres-
sure tracking is shown in Fig. 12.19. For this experiment, the engine was run with
diesel fuel. In the upper subdiagram, two injections were used to track a p r with
αv = 0.3. As αv was changed to 0.1 and Q cr was increased in the middle subdia-
gram in Fig. 12.19, the controller extrapolated (dashed) from the first detected
heat-release rate to apprehend how a post injection would affect p. The post in-
jection was then introduced, since it would decrease p r error cost, see the lower
subdiagram in Fig. 12.19.

Summary
The pressure-tracking controller was able to adjust fuel injection as p r parame-
ters were varied. The ratio between m 1f and m 2f was changed to adjust the com-
bustion duration as αv was varied. The injected fuel m f and θSOI were adjus-
ted to account for changes in Q cr and θSOC
r
, respectively. Error-free tracking could
not be obtained due to limited controllability and the steep p r increase during
constant-volume combustion. It is believed that improved tracking performance
could be obtained with a smoother p r , and an adjusted γ during the expansion

Table 12.3 Controller weights and constraints for (12.40), units for θSOI and m f
are [CAD] and [mg].

αt r = 1 R θSOI = 0.5 Rm f = 3
−0.5 ≤ ∆θSOI ≤ 0.5 −10 ≤ ∆m f ≤ 10 −25 ≤ θSOI ≤ 20 0 ≤ m f ≤ 120

243
Chapter 12. Predictive Constraint Handling and Pressure Tracking

30 1
θSOI 2
θSOI 3
θSOI
20

θ SOI [CAD]
20
θ CT [CAD]
10 0

0
−20
200 400 600 800 200 400 600 800
cycle [-] cycle [-]

150 8

p IMEP [bar]
100
m f [mg]

6
50
5

0 4
200 400 600 800 200 400 600 800
cycle [-] cycle [-]

350

300
Tex [◦ C]

250

200

150
100 200 300 400 500 600 700 800
cycle [-]

Figure 12.14 Evaluation of Tex -constraint fulfillment. The ability to fulfill Tex
was evaluated by decreasing p IMEP from 7 to 5 bar. A post-injection was intro-
duced when Tex approached Tex c = 240 ◦ C . The controller then regulated T by
ex
adjusting m 3f .

stroke. A fuel-injection system that allows for direct control of the injection rate
would improve controllability further. The controllability could also be improved
by increasing the fuel-injection rate, which would allow for an increased number
of injections in a shorter θ interval.

244
 12.5 Experimental Results

80 80

60 cycle 5 60 cycle 350

p [bar]

p [bar]
40 40

20 20

0 0
−20 0 20 40 −20 0 20 40
θ [CAD] θ [CAD]

Figure 12.15 In-cycle data for cycles 5 (left) and 350 (right) from the experiment
c = 240◦ C
in Fig. 12.14. At cycle 350, a post injection was introduced to fulfill Tex
2 r 3
whilst θCT was kept at θCT . The post-injection combustion timing θCT and cor-
responding heat-release rate dQ c3 /d θ (green) were detected by the controller.

100
p 1r

80

p 2r
60
p [bar]

40

20

−20 −10 0 10 20 30 40
θ [CAD]

Figure 12.16 The controller’s ability to follow αv changes with two injections.
During this experiment, αv was changed from 0.5 (red) to (blue) 0.2. To obtain
more constant-pressure combustion, the controller increased m 2f and delayed
1 and θ 2 .
θSOI SOI

245
Chapter 12. Predictive Constraint Handling and Pressure Tracking

100
p 2r

80

60
p 1r
p [bar]

40

20

−20 −10 0 10 20 30 40
θ [CAD]

Figure 12.17 A transition from p 1r with Q cr = 3000 J (red) to p 2r with Q cr = 6000

J (blue), with one injection. When Q cr was increased, the controller increased m f
whilst θCT was kept constant.

100

p 2r
80

60
p 1r
p [bar]

40

20

−20 −10 0 10 20 30 40
θ [CAD]

r
Figure 12.18 A θSOC transition from p 1r with θSOC
r = 10 CAD (red) to p 2r with
r r
θSOC = 5 CAD (blue). As θSOC was advanced, the controller adjusted θSOI whilst
m f was kept constant.

246
 12.5 Experimental Results

100
80
60
p [bar]

40
20
0
−20 −10 0 10 20 30 40
θ [CAD]

100
80
60
p [bar]

40
20
0
−20 −10 0 10 20 30 40
θ [CAD]

100
80
60
p [bar]

40
20
0
−20 −10 0 10 20 30 40
θ [CAD]

Figure 12.19 The suggested method for introducing injections during pressure
tracking. In the upper subdiagram, two injections were used to track a p r with
αv = 0.3. As αv was changed to 0.1 and Q cr was increased in the middle sub-
diagram, the controller extrapolated (dashed) from the first detected heat-release
rate to apprehend how a third post injection would affect p. The post-injection
was introduced since it was found to decrease p r -tracking error, see the lower
subdiagram.

247
Chapter 12. Predictive Constraint Handling and Pressure Tracking

12.6 Discussion
The use of empirical, data-based and mean-valued models to describe
in-cylinder processes has been a common theme in previous works on opti-
mal engine control for constraint fulfillment, see [Hafner et al., 2000; Stewart and
Borelli, 2008; Atkinsson et al., 2009; Karlsson et al., 2010; Grahn et al., 2014].
This chapter combined the heat-release detection and separation method
presented in Chapter 4 with the pressure-prediction method in Chapter 11. These
methods allowed for a cycle-resolved MPC formulation, where the MPC, in con-
trast to previous work, can predict the effect from different injections on the cy-
linder pressure. To the author’s knowledge, this is a novel controller-design fra-
mework that can be used for both constraint fulfillment and pressure tracking.

Constraint Fulfillment
The proposed MPC for constraint handling in (12.36) worked as intended in
the experimental evaluation. It was shown in Chapter 10 that similar transient
behavior could be obtained with a simpler PI-controller design. However, the
model-based approach adopted in this chapter allowed the controller to predict
constraint violation and in that way act beforehand. This, in combination with
constraint margins, resulted in smaller NOx overshoots, and no p max overshoots,
as compared to the PI controller in Chapter 10. Furthermore, the MPC had shor-
ter settling times of 10-20 cycles, as opposed to 40-50 cycles for the PI controller.
The model-based controller accounts for variation in combustion characteris-
tics as a function of operating point which makes the feedback loop more robust.
The centralized MPC design also allowed the controller to take into account for
cross-coupling between control of load and other engine outputs.
Compared to the heuristic PI controller in Chapter 10, the MPC framework
provided a systematic way of handling input and output constraints. All con-
straints were accounted for simultaneously, where it was straightforward to add
or remove constraints as long as meaningful solutions existed. The controller was
also flexible when adding or removing injections and adjusted the size of the op-
timization problems accordingly. The model-based methods for adding and re-
moving injections could however be developed further by, for instance, also op-
timize timings and amounts of the injections introduced.
Even though MPC has its potential benefits, it demands careful tuning so
that desired controller behavior is obtained. Poor tuning could result in patho-
logical behavior, such as control of engine load with injection timing or control
of exhaust temperature with the main-injection duration, with an offset in load
tracking as a result. Further controller development would be to also incorpo-
rate gas-system dynamics and actuators. If more accurate combustion models
are available, those could be incorporated for potential controller-performance
improvement. This may however require a more sophisticated heat-release de-
tection method.

248
 12.7 Conclusions

Combustion Detection
The combustion detection method worked well, apart from some exceptions at
cycle 60 in Fig. 12.10 and cycle 50 in Fig. 12.12. These errors occurred when a pilot
injection was introduced and the controller did not detect the pilot heat-release
rate properly. Instead, it detected the main-injection heat-release and a later
heat-release peak. One approach to further develop the heat-release detection
method would be to instead compute the likelihood of a detected peak being a
combustion timing. Such a methodology could make use of a heat-release model
and fuel-injection information.

Pressure Tracking
Although tracking of a cycle-resolved pressure reference is an unconventional
way of controlling an engine, the results presented in this chapter showed that
the presented MPC in (12.40) allowed for this approach. The convergence rate
presented here (10 to 15 engine cycles) was comparable to the results in [Zweigl
et al., 2015]. Control errors presented here were however somewhat larger than
those reported in [Jörg et al., 2015; Zweigl et al., 2015]. With a fuel-injection sys-
tems that allows for additional injections during the engine cycle, and more di-
rect control of the fuel-injection rate, it is possible that the controller presented
in this chapter could exhibit more accurate cylinder-pressure control.
If this controller design is more favorable than conventional cylinder-pressure
feedback controllers that regulate combustion-timing and indicated load re-
mains to be investigated.

12.7 Conclusions
Two model predictive controllers were presented and experimentally evalua-
ted. Both controllers utilized a linearized cylinder-pressure model and a novel
combustion-detection method in order to predict in-cylinder pressure variation
due to fuel-injection changes. Experimental results demonstrated:

• Fulfillment of constraints with respect to cylinder pressure, engine-out

NOx emissions and exhaust temperature during load changes.

• Tracking of time varying ideal-pressure-cycle trajectories.

In both cases, fuel-injections were added and removed depending on the predic-
ted in-cylinder pressure.

249
13
Conclusions and Future
Research

Demands for reduced emission levels and lowered fuel consumption have cre-
ated a need for accurate engine control. This has been illustrated in this thesis
through an investigation of how model-based closed-loop combustion control
can be used to improve the reliability of a low-emission combustion concept.
This thesis has also investigated how timings and durations of multiple injec-
tions can be decided with feedback control for efficient constraint fulfillment. In
both contexts, feedback control reduces the amount of calibration work needed
for efficient engine operation. It makes the combustion processes more robust
to changes in intake conditions, hardware aging and fuel properties, and lowers
the demands for precise actuators. The main results presented in the thesis are
summarized below, together with suggestions for future research.

PPC
This thesis investigated model-based control for improved operation of partially
premixed combustion (PPC). Designed controllers experimentally demonstrated
control of ignition delay, combustion timing and pressure-rise rate in transient
operation. Gas-exchange and fuel-injection actuations for improved low-load ef-
ficiency were also suggested.
The problems of regulating ignition delay and pressure-rise rate were studied
separately in chapters 7 and 8. Since there is an inherent trade-off between igni-
tion delay and pressure-rise rate, however, it would be interesting to investigate
concurrent control of these variables. The controller objective could be formula-
ted as a set-point tracking problem with respect to ignition-delay with an upper
bound on pressure-rise rate. This behavior could be obtained by combining the
model predictive controllers in (7.12) and (8.13).
The PPC experiments presented in the thesis were mainly limited to the
low-to-mid load operating range of the engine. This suggests that future work
should include a more detailed evaluation of controller performance at higher

250
 Chapter 13. Conclusions and Future Research

engine loads. Experiments have, however, showed that it might be difficult to cre-
ate sufficiently long ignition delays at higher loads with the current experimental
setup, and that the combustion characteristics approach those of conventional
diesel combustion at higher loads. Manente et al. [2009] suggested that long ig-
nition delays could be achieved at high load with large injections, located early
during the compression stroke (θSOI < -50 CAD). This strategy was also difficult to
implement in the experimental setup used due to preignition of such injections.
Preignition avoidance is an interesting control problem related to PPC, that has
not been discussed in this thesis. It is possible that hardware adjustments, such
as a decreased compression ratio, decreased swirl, and fuels of even higher ON
would facilitate high-load PPC operation. It is, however, also possible that such
adjustments would make low-load performance more challenging.

Optimal Control
This thesis investigated how multiple injections should be actuated for efficient
fulfillment of constraints on pressure, NOx and exhaust temperature. A hybrid
multivariate PI controller was designed and experimentally evaluated. This con-
troller showed an experimental efficiency improvement of 4-5 % compared to a
single-injection controller, as restrictive constraints on pressure and NOx were
imposed.
A predictive pressure controller was introduced, where a simple pres-
sure model was used to directly control the in-cylinder pressure using model
predictive control. This controller was capable of tracking load and efficient
combustion-timing set points, as well as fulfilling constraints. The controller was
also able to track ideal-cycle pressure curves. Further development of this con-
troller include enhancement of physical model assumptions, design of a more
robust heat-release-detection method, and further investigation of model-based
conditions for varying the number of injections. It could also be interesting to
add the common-rail pressure as a control variable for increased controllability
of the heat-release shape.
Hardware that allows for additional injections and more direct control of the
fuel-injection rate could also increase the usefulness of the controller. In-cycle
control of the cylinder pressure with the use of similar predictive methods could
also be an interesting research topic. An FPGA would be a suitable option for such
an application, as the demand for computational speed increases.

Feedforward
The controllers presented in this thesis only utilized feedback control without
any feedforward action. It is, however, believed that the performance of the
controllers presented could be improved significantly with model-based feed-
forward control, especially with respect to response times in transient operation.

251
Chapter 13. Conclusions and Future Research

Design of model-based feedforward controllers that are compatible with the pre-
sented controllers is therefore an additional suggested research topic.

Model Predictive Control

Several of the controllers presented in the thesis were based on the principle of
model predictive control (MPC). MPC provides a model-based framework that
systematically handles output constraints, and the increasing complexity of en-
gine control, for which the number of sensors and actuators have increased du-
ring the last decades. MPC also conveniently handled the problem of varying the
number of inputs in Chapter 12, where the number of injections changed depen-
ding on the engine operating point.
The approach taken in this work was to repeatedly linearize the system mo-
del and solve a quadratic program. An alternative that would require additional
memory storage but less online computations is explicit MPC, where the optimi-
zation problem is solved offline and the optimal input signal is stored in a lookup
table [Bemporad et al., 2002]. It is also possible that nonlinear solvers will be fast
enough in the near future for the original nonlinear optimization problems to be
solved online. Moreover, the problem of deciding the number of injections is a
discrete optimization problem. This suggests that a hybrid MPC solution could
be used to solve the control problems studied in chapters 10 and 12. These three
approaches to MPC are interesting extensions to the work presented here and
deserve future investigation.
When solving constrained optimization problems, such as in MPC, one ob-
tains Lagrangian multipliers. These can be interpreted as costs with respect to
reference-tracking deviation or efficiency, that has to be paid to fulfill the diffe-
rent constraints. A suggestion for future work is to incorporate these multipliers
in engine diagnostics, to analyze and adjust controller constraints so that the
price of constraint fulfillment does not become too large. If for instance, it is no-
ted that fulfillment of constraints on pressure-rise rate or NOx emissions requi-
res a late and inefficient combustion timing at a certain operating point. Then,
a supervisory controller could take action by increasing the EGR set point, or by
adjusting the efficiency of the after-treatment system to increase the overall en-
gine efficiency.
An additional suggestion for future work would be to investigate how the
MPC formulations presented here could be extended or adjusted to find effi-
cient compromises between conflicting constraints such as emissions of soot
and NOx , or ignition delay and pressure-rise rate.

252
Bibliography

A, B. Kempinski, and J. Rife (1981). Knock in Spark Ignition Engines. SAE Technical
Paper 810147.
Agrell, F., H.-E. Ångström, B. Eriksson, J. Wikander, and J. Linderyd (2003). In-
tegrated Simulation and Engine Test of Closed Loop HCCI Control by aid of
Variable Valve Timings. SAE Technical Paper 2003-01-0748.
Akihama, K., Y. Takatori, K. Inagaki, S. Sasaki, and A. M. Dean (2001). Mechanism
of the Smokeless Rich Diesel Combustion by Reducing Temperature. SAE Te-
chnical Paper 2001-01-0655.
Annand, W. (1963). “Heat transfer in the cylinders of reciprocating internal com-
bustion engines”. Proc. Inst. of Mech. Eng. 177:1, pp. 973–996.
Aoyagi, Y., T. Kamimoto, Y. Matsui, and S. Matsuoka (1980). A Gas Sampling Study
on the Formation Processes of Soot and NO in a DI Diesel Engine. SAE Techni-
cal Paper 800254.
Askin, A. C., G. E. Barter, T. H. West, and D. K. Manley (2015). “The heavy-duty
vehicle future in the United States: a parametric analysis of technology and
policy tradeoffs”. Energy Policy 81, pp. 1–13.
Åström, K. J. and T. Hägglund (2006). Advanced PID Control. ISA – The Instrumen-
tation, Systems, and Automation Society, Research Triangle Park, NC, USA.
Åström, K. J. and R. M. Murray (2010). Feedback Systems: an Introduction for
Scientists and Engineers. Princeton University Press, Princeton, NJ, USA.
Atkinson, C. M., M. Allain, Y. Kalish, and H. Zhang (2009). Model-Based Con-
trol of Diesel Engines for Fuel Efficiency Optimization. SAE Technical Paper
2009-01-0727.
Bemporad, A., M. Morari, V. Dua, and E. N. Pistikopoulos (2002). “The explicit
linear quadratic regulator for constrained systems”. Automatica 38:1, pp. 3–
20.

253
Bibliography

Bengtsson, J., P. Strandh, R. Johansson, P. Tunestål, and B. Johansson (2004).

“Closed-loop combustion control of homogeneous charge compression igni-
tion (HCCI) engine dynamics”. Int. J. Adaptive Control and Signal Processing
18:2, pp. 167–179.
Bengtsson, J., P. Strandh, R. Johansson, P. Tunestål, and B. Johansson (2006). “Mo-
del predictive control of homogeneous charge compression ignition (HCCI)
engine dynamics”. In: Proc. IEEE Int. Conf. Control Applications. Münich,
Germany, pp. 1675–1680.
Blom, D., M. Karlsson, K. Ekholm, P. Tunestål, and R. Johansson (2008). HCCI En-
gine Modeling and Control using Conservation Principles. SAE Technical Pa-
per 2008-01-0789.
Borgqvist, P. (2013). The Low Load Limit of Gasoline Partially Premixed Combus-
tion (PPC)-Experiments in a Light Duty Diesel Engine. TMHP-13/1091. PhD
thesis. Dept. Energy Sciences, Lund University, Lund, Sweden.
Borgqvist, P., P. Tunestål, and B. Johansson (2012). Gasoline Partially Premixed
Combustion in a Light Duty Engine at Low Load and Idle Operating Condi-
tions. SAE Technical Paper 2012-01-0687.
Borgqvist, P., P. Tunestål, and B. Johansson (2013). “Comparison of negative valve
overlap (NVO) and rebreathing valve strategies on a gasoline PPC engine at
low load and idle operating conditions”. SAE Int. J. Engines 6:2013-01-0902,
pp. 366–378.
BorgWarner (24, 2014). Borgwarner equips new generation of diesel engines from
Volkswagen with pressure sensor glow plugs. press release. BorgWarner. URL:
https : / / cdn . borgwarner . com / docs / default - source / press -
release - downloads / en _ bw - equips - new - generation - of - diesel -
engines-from-vw-with-psg.pdf (visited on 2018).
Bosch, R. G. (2011). Bosch Automotive Handbook, 8th Edition. Wiley, Chichester,
West Sussex, England.
Broy, M., I. H. Kruger, A. Pretschner, and C. Salzmann (2007). “Engineering auto-
motive software”. Proc. IEEE 95:2, pp. 356–373.
Brunt, M. F. and C. R. Pond (1997). Evaluation of Techniques for Absolute Cylinder
Pressure Correction. SAE Technical Paper 970036.
Casella, G. and R. L. Berger (2002). Statistical Inference. Duxbury, Pacific Grove,
CA, USA.
Castellano, J., A. Chaudhari, and J. Bromham (2013). Adaptive Temperature
Control for Diesel Particulate Filter Regeneration. SAE Technical Paper
2013-01-0517.
Chartier, C., O. Andersson, B. Johansson, M. Musculus, and M. Bobba (2011). “Ef-
fects of post-injection strategies on near-injector over-lean mixtures and un-
burned hydrocarbon emission in a heavy-duty optical diesel engine”. SAE Int.
J. Engines 4:1, pp. 1978–1992.

254
 Bibliography

Chiang, C. and A. Stefanopoulou (2005). “Control of thermal ignition delay in

gasoline engines”. In: Proc. American Control Conf. (ACC 2005). Cape Town,
South Africa, pp. 3847–3852.
Chiang, C. and A. Stefanopoulou (2006). “Sensitivity analysis of combustion ti-
ming and duration of homogeneous charge compression ignition (HCCI) en-
gines”. In: Proc. American Control Conf. (ACC 2006). Minneapolis, MN, USA,
pp. 6–12.
Chmela, F. G., G. H. Pirker, and A. Wimmer (2007). “Zero-dimensional ROHR si-
mulation for DI diesel engines–a generic approach”. Energy Conversion and
Management 48:11, pp. 2942–2950.
Cho, E., K. Thoney, T. Hodgson, and R. King (2003). “Rolling horizon scheduling
of multi-factory supply chains”. In: Proc. 2003 Winter Simulation Conf. New
Orleans, LA, USA, pp. 1409–1416.
Contestabile, M., G. Offer, R. Slade, F. Jaeger, and M. Thoennes (2011). “Battery
electric vehicles, hydrogen fuel cells and biofuels. which will be the winner?”
Energy & Environmental Science 4:10, pp. 3754–3772.
Curran, H. J., P. Gaffuri, W. Pitz, and C. Westbrook (2002). “A comprehensive mo-
deling study of iso-octane oxidation”. Combustion and Flame 129:3, pp. 253–
280.
Dawid, H. (2005). “Long horizon versus short horizon planning in dynamic op-
timization problems with incomplete information”. Economic Theory 25:3,
pp. 575–597.
Dec, J. E. (1997). A Conceptual Model of DI Diesel Combustion based on
Laser-Sheet Imaging. SAE Technical Paper 970873.
DelVescovo, D., S. Kokjohn, and R. Reitz (2016). “The development of an igni-
tion delay correlation for PRF fuel blends from PRF0 (n-heptane) to PRF100
(iso-octane)”. SAE Int. J. Engines 9:1, pp. 520–535.
des Constructeurs d’Automobiles, Organisation Internationale (2017). Produc-
tion statistics. URL: http : / / http : / / www . oica . net / category /
production-statistics/2017-statistics/ (visited on 2018).
Dober, G., S. Tullis, G. Greeves, N. Milovanovic, M. Hardy, and S. Zuelch (2008).
The Impact of Injection Strategies on Emissions Reduction and Power Output
of Future Diesel Engines. SAE Technical Paper 2008-01-0941.
Douaud, A. and P. Eyzat (1978). Four-Octane-Number Method for Predicting the
Anti-Knock Behavior of Fuels and Engines. SAE Technical Paper 780080.
Egnell, R. (2001). On Zero-Dimensional Modelling of Combustion and NOx for-
mation in diesel engines. PhD thesis. Dept. Heat and Power Engineering, Lund
University, Lund, Sweden.

255
Bibliography

Eichmeier, J., U. Wagner, and U. Spicher (2012). “Controlling gasoline low tem-
perature combustion by diesel micro pilot injection”. J. Eng. for Gas Turbines
and Power 134:7.
Ekholm, K., M. Karlsson, P. Tunestål, R. Johansson, B. Johansson, and P. Strandh
(2008a). “Ethanol-diesel fumigation in a multi-cylinder engine”. SAE Int. J.
Fuels and Lubricants 1, pp. 26–36.
Ekholm, K., M. Karlsson, P. Tunestål, R. Johansson, B. Johansson, and P. Strandh
(2008b). “Ethanol-diesel fumigation in a multi-cylinder engine”. SAE Int. J.
Fuels Lubricants 1, pp. 26–36.
Enkvist, P., T. Nauclér, and J. Rosander (2007). “A cost curve for greenhouse gas
reduction”. McKinsey Quarterly 1, p. 34.
Epping, K., S. Aceves, R. Bechtold, and J. E. Dec (2002). The Potential of HCCI
Combustion for High Efficiency and Low Emissions. SAE Technical Paper
2002-01-1923.
Eriksson, L. (1998). Requirements for and a Systematic Method for Identifying
Heat-Release Model Parameters. SAE Technical Paper 980626.
Eriksson, L. and L. Nielsen (2014). Modeling and Control of Engines and Driveli-
nes. John Wiley & Sons, Chichester, West Sussex, England.
Eriksson, L. and M. Sivertsson (2016). “Calculation of optimal heat release rates
under constrained conditions”. SAE Int. J. Engines 9:2, pp. 1143–1162.
Eriksson, L. and A. Thomasson (2017). “Cylinder state estimation from measu-
red cylinder pressure traces-a survey”. IFAC-PapersOnLine 50:1, pp. 11029–
11039.
European Commission (2016). A European Strategy for Low-Emission Mobility.
COM(2016) 501. Brussels, Belgium.
European Union (2007). “Cleaner trucks and buses: tighter limits for nitrogen oxi-
des and particulate matter (Euro VI).” Press release, IP/07/1989.
Fenimore, C. (1971). “Formation of nitric oxide in premixed hydrocarbon fla-
mes”. In: Proc. Int. Symposium on Combustion. Vol. 13. 1. Pittsburgh, PA,
pp. 373–380.
Fieweger, K., R. Blumenthal, and G. Adomeit (1997). “Self-ignition of SI engine
model fuels: a shock tube investigation at high pressure”. Combustion and
Flame 109:4, pp. 599–619.
Fiveland, S. B. and D. N. Assanis (2001). Development of a two-zone HCCI com-
bustion model accounting for boundary layer effects. SAE Technical Paper
2001-01-1028.
Fridriksson, H., B. Sundén, S. Hajireza, and M. Tunér (2011). CFD Investigation of
Heat Transfer in a Diesel Engine with Diesel and PPC Combustion Modes. SAE
Technical Paper 2011-01-1838.

256
 Bibliography

Garpinger, O. and T. Hägglund (2015). “Software-based optimal PID design with

robustness and noise sensitivity constraints”. J. Process Control 33, pp. 90–
101.
Gatowski, J., E. N. Balles, K. Chun, F. Nelson, J. Ekchian, and J. B. Heywood (1984).
Heat Release Analysis of Engine Pressure Data. SAE Technical Paper 841359.
Geisser, S. (1993). Predictive Inference. Chapman & Hall, New York, NY, USA.
Gieshoff, J., A. Schäfer-Sindlinger, P. Spurk, J. Van Den Tillaart, and G. Garr (2000).
Improved SCR Systems for Heavy Duty Applications. SAE Technical Paper
2000-01-0189.
Gong, J. and C. Rutland (2013). A Quasi-Dimensional NOx Emission Model for
Spark Ignition Direct Injection (SIDI) Gasoline Engines. SAE Technical Paper
2013-01-1311.
Gordon, N. J., D. J. Salmond, and A. F. Smith (1993). “Novel approach to
nonlinear/non-Gaussian Bayesian state estimation”. In: IEE Proc. F Radar
and Signal Processing. Vol. 140. 2, pp. 107–113.
Grahn, M., K. Johansson, and T. McKelvey (2014). “A transient diesel EMS strategy
for online implementation”. In: Proc. IFAC Volumes. Vol. 47. 3. Cape Town,
South Africa, pp. 11842–11847.
Guardiola, C., B. Pla, A. Garcia, and V. Boronat (2017). “Optimal heat release sha-
ping in a reactivity controlled compression ignition (RCCI) engine”. Control
Theory and Technology 15:2, pp. 117–128.
Guzzella, L. and C. Onder (2009). Introduction to Modeling and Control of Inter-
nal Combustion Engine Systems. Springer-Verlag, Berlin, Germany.
Haagen-Smit, A. (1952). “Chemistry and physiology of Los Angeles smog”. Indus-
trial & Eng. Chemistry 44:6, pp. 1342–1346.
Hafner, M., M. Schüler, O. Nelles, and R. Isermann (2000). “Fast neural networks
for diesel engine control design”. Control Eng. Practice 8:11, pp. 1211–1221.
Halstead, M., L. Kirsch, A. Prothero, and C. Quinn (1975). “A mathematical mo-
del for hydrocarbon autoignition at high pressures”. In: Proc. Royal Society of
London A: Mathematical, Physical and Eng. Sciences. Vol. 346. 1647, pp. 515–
538.
Halstead, M., L. Kirsch, and C. Quinn (1977). “The autoignition of hydrocarbon
fuels at high temperatures and pressures—Fitting of a mathematical model”.
Combustion and Flame 30, pp. 45–60.
Han, J.-S., P.-H. Lu, X.-B. Xie, M.-C. Lai, and N. A. Henein (2002). Investigation of
Diesel Spray Primary Break-Up and Development for Different Nozzle Geome-
tries. SAE Technical Paper 2002-01-2775.
Han, Z., A. Uludogan, G. J. Hampson, and R. D. Reitz (1996). Mechanism of Soot
and NOx Emission Reduction using Multiple-Injection in a Diesel Engine. SAE
Technical Paper 960633.

257
Bibliography

Hanson, R. M., S. L. Kokjohn, D. A. Splitter, and R. D. Reitz (2010). “An ex-

perimental investigation of fuel reactivity controlled PCCI combustion in a
heavy-duty engine”. SAE Int. J. Engines 3:1, pp. 700–716.
Haraldsson, G. (2005). Closed-loop Combustion Control of a Multi Cylinder HCCI
Engine using Variable Compression Ratio and Fast Thermal Management.
TMHP–05/1028. PhD thesis. Dept. of Heat and Power Engineering, Lund
University, Lund, Sweden.
Haraldsson, G., P. Tunestål, B. Johansson, and J. Hyvönen (2002). HCCI Combus-
tion Phasing in a Multi Cylinder Engine using Variable Compression Ratio.
SAE Technical Paper 2002-01-2858.
Hasegawa, R. and H. Yanagihara (2003). HCCI Combustion in DI Diesel Engine.
SAE Technical Paper 2003-01-0745.
Hast, M., K. Åström, B. Bernhardsson, and S. Boyd (2013). “PID design by
convex-concave optimization”. In: Proc. European Control Conf. (ECC 2013).
Zürich, Switzerland, pp. 4460–4465.
Hellström, E., D. Lee, L. Jiang, A. G. Stefanopoulou, and H. Yilmaz (2013).
“On-board calibration of spark timing by extremum seeking for flex-fuel en-
gines”. IEEE Trans. Control Systems Technology 21:6, pp. 2273–2279.
Hendricks, E. (1986). A Compact, Comprehensive Model of Large Turbocharged,
Two-Stroke Diesel Engines. SAE Technical Paper 861190.
Henningsson, M. (2012). Data-Rich Multivariable Control of Heavy-Duty Engi-
nes. PhD thesis TFRT-1092. Dept. Automatic Control, Lund University, Lund,
Sweden.
Henningsson, M., K. Ekholm, P. Strandh, P. Tunestål, and R. Johansson (2012).
“Dynamic mapping of diesel engine through system identification”. In: D. Al-
berer H. Hjalmarsson, L. (Ed.). Identification for Automotive Systems. Sprin-
ger, London, England, pp. 223–239.
Heywood, J. B. (1988). Internal Combustion Engine Fundamentals. McGraw-Hill,
New York, NY, USA.
Hildingsson, L., G. Kalghatgi, N. Tait, B. Johansson, and A. Harrison (2009). Fuel
Octane Effects in the Partially Premixed Combustion Regime in Compression
Ignition Engines. SAE Technical Paper 2009-01-2648.
Hohenberg, G. F. (1979). Advanced Approaches for Heat Transfer Calculations.
SAE Technical Paper 790825.
Honardar, S., H. Busch, T. Schnorbus, C. Severin, A. F. Kolbeck, and T. Korfer
(2011). “Exhaust temperature management for diesel engines assessment of
engine concepts and calibration strategies with regard to fuel penalty”. In:
Proc. 10th Int. Conf. Engines Vehicles.
Iorio, B., V. Gioglio, G. Police, and N. Rispoli (2003). Methods of Pressure Cycle
Processing for Engine Control. SAE Technical Paper 2003-01-0352.

258
 Bibliography

IPCC (2014). “Climate change 2014, synthesis report”. IPCC Fifth Assessment Re-
port.
Isermann, R. (2014). Engine Modeling and Control. Springers Berlin Heidelberg,
Berlin, Germany.
Johansson, B. (2006). Förbränningsmotorer. Institutionen för värme-och kraftte-
knik, Lunds Tekniska Högskola, Lund, Sweden.
Johansson, R. (1993). System Modeling and Identification. Prentice Hall, En-
glewood Cliffs, NJ, USA.
Johansson, T. B., P. Kågesson, H. Johansson, L. Jonsson, J. Westin, H. Hejenstedt,
O. Hådell, K. Holmgren, and P. Wollin (2013). “Fossilfrihet på väg”. Ministry of
Enterprise, SOU 2013:84. Stockholm, Sweden.
Jones, J. C. P. and J. Frey (2015). “Threshold optimization and performance eval-
uation of a classical knock controller”. SAE Int. J. Engines 8:3, pp. 1021–1028.
Jörg, C., T. Schnorbus, S. Jarvis, B. Neaves, K. Bandila, and D. Neumann (2015).
“Feedforward control approach for digital combustion rate shaping reali-
zing predefined combustion processes”. SAE Int. J. Engines 8:2015-01-0876,
pp. 1041–1054.
Julier, S. J. and J. K. Uhlmann (2004). “Unscented filtering and nonlinear estima-
tion”. Proc. IEEE 92:3, pp. 401–422.
Jury, E. I. (1964). “Theory and application of the Z-transform method”.
Kalghatgi, G. T., P. Risberg, and H.-E. Ångström (2006). Advantages of Fuels with
High Resistance to Auto-Ignition in Late-Injection, Low-Temperature, Com-
pression Ignition Combustion. SAE Technical Paper 2006-01-3385.
Källkvist, K. (2011). Fuel Pressure Modelling in a Common-Rail Direct Injection
System. MSc Thesis LiTH-ISY-EX–11/4488–SE. Linköping University, Dept. of
Electrical Engineering, Linköping, Sweden.
Kalman, R. (1959). “On the general theory of control systems”. IRE Trans. Auto-
matic Control 4:3, pp. 110–110.
Kalman, R. E. et al. (1960). “Contributions to the theory of optimal control”. Bol.
Soc. Mat. Mexicana 5:2, pp. 102–119.
Kalman, R. E. (1960). “A new approach to linear filtering and prediction pro-
blems”. J. Fluids Eng. 82:1, pp. 35–45.
Kamimoto, T. and M.-h. Bae (1988). High Combustion Temperature for the Reduc-
tion of Particulate in Diesel Engines. SAE Technical Paper 880423.
Kanda, T., T. Hakozaki, T. Uchimoto, J. Hatano, N. Kitayama, and H. Sono (2005).
PCCI Operation with Early Injection of Conventional Diesel Fuel. SAE Techni-
cal Paper 2005-01-0378.
Karlsson, M., K. Ekholm, P. Strandh, R. Johansson, and P. Tunestål (2008). “LQG
control for minimization of emissions in a diesel engine”. In: Proc. IEEE Int.
Conf. Control Applications (CCA 2008). San Antonio, TX, USA, pp. 245–250.

259
Bibliography

Karlsson, M., K. Ekholm, P. Strandh, R. Johansson, and P. Tunestål (2010).

“Multiple-input multiple-output model predictive control of a diesel engine”.
In: Proc. 6th IFAC Symposium Advances in Automotive Control (AAC 2010).
Münich, Germany, pp. 131–136.
Katare, S. R. and P. M. Laing (2009). “Hydrogen in diesel exhaust: effect on diesel
oxidation catalyst flow reactor experiments and model predictions”. SAE Int.
J. Fuels and Lubricants 2:1, pp. 605–611.
Kay, S. M. (1998). Fundamentals of Statistical Signal Processing, Vol. II: Detection
Theory. Prentice Hall, Upper Saddle River, NJ, USA.
Kiencke, U. and L. Nielsen (2000). Automotive Control Systems. For Engine, Drive-
line, and Vehicle. Springer-Verlag, Berlin.
Killingsworth, N. J., S. M. Aceves, D. L. Flowers, F. Espinosa-Loza, and M. Krstić
(2009). “HCCI engine combustion-timing control: optimizing gains and fuel
consumption via extremum seeking”. IEEE Trans. Control Systems Technology
17:6, pp. 1350–1361.
Kim, D., I. Ekoto, W. F. Colban, and P. C. Miles (2008). In-Cylinder CO and UHC
Imaging in a Light-Duty Diesel Engine During PPCI Low-Temperature Com-
bustion. SAE Technical Paper 2008-01-1602.
Kimura, S., O. Aoki, H. Ogawa, S. Muranaka, and Y. Enomoto (1999). New Com-
bustion Concept for Ultra-Clean and High-Efficiency Small DI Diesel Engines.
SAE Technical Paper 1999-01-3681.
Kitamura, T., T. Ito, J. Senda, and H. Fujimoto (2002). “Mechanism of smokeless
diesel combustion with oxygenated fuels based on the dependence of the
equivalence ratio and temperature on soot particle formation”. Int. J. Engine
Research 3:4, pp. 223–248.
Klein, M. (2007). Single-Zone Cylinder Pressure Modeling and Estimation for Heat
Release Analysis of SI Engines. Dissertation No. 1124, Department of Electrical
Engineering. PhD thesis. Linköping University, Linköping, Sweden.
Klein, M., L. Eriksson, and J. Åslund (2006). “Compression ratio estimation based
on cylinder pressure data”. Control Eng. Practice 14:3, pp. 197–211.
Kokjohn, S. L., R. M. Hanson, D. Splitter, and R. Reitz (2011). “Fuel reactivity con-
trolled compression ignition (RCCI): a pathway to controlled high-efficiency
clean combustion”. Int. J. Engine Research 12:3, pp. 209–226.
Kokjohn, S. L., R. M. Hanson, D. A. Splitter, and R. D. Reitz (2009). Experiments
and Modeling of Dual-Fuel HCCI and PCCI Combustion using in-Cylinder
Fuel Blending. SAE Technical Paper 2009-01-2647.
Kook, S., C. Bae, P. C. Miles, D. Choi, and L. M. Pickett (2005). The Influence of
Charge Dilution and Injection Timing on Low-Temperature Diesel Combus-
tion and Emissions. SAE Technical Paper 2005-01-3837.

260
 Bibliography

Krieger, R. and G. Borman (1966). “The computation of apparent heat release in

IC engines”. ASME Series 66.
Lee, H., Y. Park, S. Hong, M. Lee, and M. Sunwoo (2013). EGR Rate Estimation
for Cylinder Air Charge in a Turbocharged Diesel Engine using an Adaptive
Observer. SAE Technical Paper 2013-01-0246.
Leen, G. and D. Heffernan (2002). “Expanding automotive electronic systems”.
Computer 35:1, pp. 88–93.
Lewander, M., B. Johansson, P. Tunestål, N. Keeler, N. Milovanovic, and P. Bergs-
trand (2008). “Closed loop control of a partially premixed combustion engine
using model predictive control strategies”. In: Proc. of AVEC. Vol. 8. Kobe, Ja-
pan.
Lewander, M., A. Widd, B. Johansson, and P. Tunestål (2012). “Steady state fuel
consumption optimization through feedback control of estimated cylinder
individual efficiency”. In: Proc. American Control Conf. (ACC 2012). Fairmont
Queen Elizabeth, Montreal, Canada, pp. 4210–4214.
Li, C., L. Yin, S. Shamun, M. Tunér, B. Johansson, R. Solsjö, and X.-S. Bai (2016).
“Transition from HCCI to PPC: the sensitivity of combustion phasing to the
intake temperature and the injection timing with and without EGR”. SAE Te-
chnical Paper 0767.
Livengood, J. and P. Wu (1955). “Correlation of autoignition phenomena in in-
ternal combustion engines and rapid compression machines”. In: Proc. Int.
Symposium Combustion. Vol. 5. 1. Reinhold, NY, USA, pp. 347–356.
Lönn, S., A. Matamis, M. Tunér, M. Richter, and Ö. Andersson (2017). Optical
Study of Fuel Spray Penetration and Initial Combustion Location under PPC
Conditions. SAE Technical Paper 2017-01-0752.
Lu, T. and C. K. Law (2009). “Toward accommodating realistic fuel chemistry in
large-scale computations”. Progress in Energy and Combustion Science 35:2,
pp. 192–215.
Maciejowski, J. M. (2002). Predictive Control: with Constraints. Pearson Educa-
tion, Essex, England.
MacMillan, D., A. La Rocca, P. Shayler, T. Morris, M. Murphy, and I. Pegg (2009).
“Investigating the effects of multiple injections on stability at cold idle for a
DI diesel engine”. SAE Int. J. Engines 2:1.
Majewski, W. A. and M. K. Khair (2006). Diesel Emissions and their Control. SAE
International, Warrendale, PA, USA.
Manente, V., B. Johansson, and P. Tunestål (2009). Partially Premixed Combus-
tion at High Load using Gasoline and Ethanol, a Comparison with Diesel. SAE
Technical Paper 2009-01-0944.

261
Bibliography

Manente, V., B. Johansson, and P. Tunestål (2010a). “Characterization of partially

premixed combustion with ethanol: EGR sweeps, low and maximum loads”.
J. Eng. for Gas Turbines and Power-Transactions ASME 132:8.
Manente, V., P. Tunestål, B. Johansson, and W. J. Cannella (2010b). Effects of Etha-
nol and Different Type of Gasoline Fuels on Partially Premixed Combustion
from Low to High Load. SAE Technical Paper 2010-01-0871.
Manente, V., A. Vressner, P. Tunestål, and B. Johansson (2008). Validation of a Self
Tuning Gross Heat Release Algorithm. SAE Technical Paper 2008-01-1672.
Manente, V., C.-G. Zander, B. Johansson, P. Tunestål, and W. Cannella (2010c). An
Advanced Internal Combustion Engine Concept for Low Emissions and High
Efficiency from Idle to Max Load using Gasoline Partially Premixed Combus-
tion. SAE Technical Paper 2010-01-2198.
Martinez-Frias, J., S. M. Aceves, D. Flowers, J. R. Smith, and R. Dibble (2000). HCCI
Engine Control by Thermal Management. SAE Technical Paper 2000-01-2869.
McKinsey & Company (2014). Electric Vehicles in Europe: Gearing up for a new
phase. Technical Report.
Meyer, J. (2011). Calibration reduction in internal combustion engine fueling con-
trol: modeling, estimation and stability robustness. PhD thesis. The Ohio State
University.
Miller, J. A. and C. T. Bowman (1989). “Mechanism and modeling of nitrogen
chemistry in combustion”. Progress in Energy and Combustion Science 15:4,
pp. 287–338.
Murić, K., O. Stenlåås, and P. Tunestål (2014). “Zero-dimensional modeling of
NOx formation with least squares interpolation”. Int. J. Engine Research 15:8,
pp. 944–953.
Musculus, M. P. (2006). Multiple Simultaneous Optical Diagnostic Imaging of
Early-Injection Low-Temperature Combustion in a Heavy-Duty Diesel Engine.
SAE Technical Paper 2006-01-0079.
Nagatsu, K., A. Inoue, K. Matsumoto, T. Kaminaga, T. Miyamoto, and T. Youso
(2017). Control device for compression ignition-type engine. US Patent
9,719,441.
Nilsson, Y. and L. Eriksson (2001). “A new formulation of multi-zone combustion
engine models”. IFAC Proc. Volumes 34:1, pp. 357–362.
Noehre, C., M. Andersson, B. Johansson, and A. Hultqvist (2006). Characteriza-
tion of Partially Premixed Combustion. SAE Technical Paper 2006-01-3412.
Okamoto, T. and N. Uchida (2016). “New concept for overcoming the trade-off
between thermal efficiency, each loss and exhaust emissions in a heavy duty
diesel engine”. SAE Int. J. Engines 9:2, pp. 859–867.
Olsson, J.-O., P. Tunestål, G. Haraldsson, and B. Johansson (2001). “A turbochar-
ged dual-fuel HCCI engine”. SAE Special Publications 2001:1627.

262
 Bibliography

Olsson, J.-O., P. Tunestål, and B. Johansson (2004). Boosting for High Load HCCI.
SAE Technical Paper 2004-01-0940.
Onishi, S., S. H. Jo, K. Shoda, P. Do Jo, and S. Kato (1979). Active
Thermo-Atmosphere Combustion (ATAC)-a New Combustion Process for
Internal Combustion Engines. SAE Technical paper, 790501.
Osuka, I., M. Nishimura, Y. Tanaka, and M. Miyaki (1994). Benefits of new Fuel In-
jection System Technology on Cold Startability of Diesel Engines-Improvement
of Cold Startability and White Smoke Reduction by Means of Multi Injection
with Common Rail Fuel System ECD-U2. SAE Technical Paper 940586.
Ott, T., F. Zurbriggen, C. H. Onder, and L. Guzzella (2013). “Cylinder individual
feedback control of combustion in a dual fuel engine”. In: Advances in Auto-
motive Control (ACC 2013). Vol. 7. 1. Tokyo, Japan, pp. 600–605.
Powell, J. D. (1993). “Engine control using cylinder pressure: past, present, and
future”. J. Dynamic Systems, Measurements, and Control 115:2B, pp. 343–350.
Randolph, A. L. (1990). Methods of Processing Cylinder-Pressure Transducer Sig-
nals to Maximize Data Accuracy. SAE Technical Paper 900170.
Rassweiler, G. M. and L. Withrow (1938). Motion Pictures of Engine Flames Corre-
lated with Pressure Cards. SAE Technical Paper 380139.
Roelle, M. J., N. Ravi, A. F. Jungkunz, and J. C. Gerdes (2006). “A dynamic model of
recompression HCCI combustion including cylinder wall temperature”. In:
Proc. IMECE. Chicago, IL, USA, pp. 415–424.
Saracino, R., M. R. Gaballo, S. Mannal, S. Motz, A. Carlucci, and M. Benegiamo
(2015). Cylinder Pressure-Based Closed Loop Combustion Control: A Valid
Support to Fulfill Current and Future Requirements of Diesel Powertrain Sys-
tems. SAE Technical Paper 2015-24-2423.
Shaver, G. M., J. C. Gerdes, and M. Roelle (2004). “Physics-based closed-loop con-
trol of phasing, peak pressure and work output in HCCI engines utilizing va-
riable valve actuation”. In: Proc. American Control Conf. (ACC 2004). Vol. 1.
Boston, MA, USA, pp. 150–155.
Sher, E. (1998). Handbook of Air Pollution from Internal Combustion Engines: Pol-
lution Formation and Control. Academic Press, San Diego, CA, USA.
Shi, Y., H.-W. Ge, and R. D. Reitz (2011). Computational Optimization of Internal
Combustion Engines. Springer Science & Business Media, London, England.
Sjöberg, M. and J. E. Dec (2005). “An investigation into lowest acceptable com-
bustion temperatures for hydrocarbon fuels in HCCI engines”. Proc. Combus-
tion Institute 30:2, pp. 2719–2726.
Solaka, H., U. Aronsson, M. Tunér, and B. Johansson (2012). Investigation of Par-
tially Premixed Combustion Characteristics in Low Load Range with Regards
to Fuel Octane Number in a Light-Duty Diesel Engine. SAE Technical Paper
2012-01-0684.

263
Bibliography

Solsjö, R. (2014). Large Eddy Simulation of Turbulent Combustion in PPC and Die-
sel Engines. PhD thesis TMHP-14/1107, Dept. Energy Sciences. PhD thesis.
Lund University, Lund, Sweden.
Spadaccini, L. J. and J. A. Tevelde (1982). “Autoignition characteristics of
aircraft-type fuels”. Combustion and Flame 46, pp. 283–300.
Sperling, D. and D. Gordon (2009). Two Billion Cars: Driving Toward Sustainabi-
lity. Oxford University Press, New York, NY, USA.
Stas, M. J. (1996). Thermodynamic Determination of TDC in Piston Combustion
Engines. SAE Technical Paper 960610.
Stewart, G. and F. Borrelli (2008). “A model predictive control framework for in-
dustrial turbodiesel engine control”. In: Proc. Conf. Decision and Control.
Cancun, Mexico, pp. 5704–5711.
Szekely, G. A., A. S. Solomon, and P.-H. Tsai (2004). Optimization of the
Stratified-Charge Regime of the Reverse-Tumble Wall-Controlled Gasoline
Direct-Injection Engine. SAE Technical Paper 2004-01-0037.
Takeda, Y. and N. Keiichi (1996). Emission Characteristics of Premixed Lean Diesel
Combustion with Extremely Early Staged Fuel Injection. SAE Technical Paper
961163.
Tanov, S., R. Collin, B. Johansson, and M. Tunér (2014). “Combustion stratifica-
tion with partially premixed combustion, PPC, using NVO and split injection
in a LD-diesel engine”. SAE Int. J. Engines 7:4, pp. 1911–1919.
Tsurushima, T., E. Kunishima, Y. Asaumi, Y. Aoyagi, and Y. Enomoto (2002). The
Effect of Knock on Heat Loss in Homogeneous Charge Compression Ignition
Engines. SAE Technical Paper 2002-01-0108.
Tunestål, P. (2011). “TDC offset estimation from motored cylinder pressure data
based on heat release shaping”. Oil & Gas Science and Technology–Revue
d’IFP Energies nouvelles 66:4, pp. 705–716.
Tunestål, P., J. K. Hedrick, and R. Johansson (2001). “Model-based estimation of
cylinder pressure sensor offset using least-squares methods”. In: Proc. Conf.
Decision and Control. Vol. 4. Orlando, FL, USA, pp. 3740–3745.
Turns, S. R. et al. (1996). An Introduction to Combustion. McGraw-Hill, New York,
NY, USA.
Wahlström, J. and L. Eriksson (2011). “Modelling diesel engines with a
variable-geometry turbocharger and exhaust gas recirculation by optimi-
zation of model parameters for capturing non-linear system dynamics”.
Proc. Inst. Mech. Eng., Part D: J. Automobile Eng. 225:7, pp. 960–986.
Wang, Y. and S. Boyd (2010). “Fast model predictive control using online optimi-
zation”. IEEE Trans. Control Systems Technology 18:2, pp. 267–278.

264
 Bibliography

Weall, A. and N. Collings (2009). Gasoline Fuelled Partially Premixed Compression

Ignition in a Light Duty Multi Cylinder Engine: a Study of Low Load and Low
Speed Operation. SAE Technical Paper 2009-01-1791.
Westbrook, C. K. and F. L. Dryer (1981). “Simplified reaction mechanisms for the
oxidation of hydrocarbon fuels in flames”. Combustion Science and Techno-
logy 27:1-2, pp. 31–43.
Widd, A., K. Ekholm, P. Tunestål, and R. Johansson (2009). “Experimental evalua-
tion of predictive combustion phasing control in an HCCI engine using fast
thermal management and VVA”. In: Proc. Control Applications & Intelligent
Control. Saint Petersburg, Russia, pp. 334–339. DOI: 10 . 1109 / CCA . 2009 .
5281087.
Widd, A., P. Tunestål, J. Åkesson, and R. Johansson (2012). “Single-zone diesel PPC
modeling for control”. In: Proc. American Control Conf. (ACC 2012). Montréal,
Canada, pp. 5731–5736.
Widd, A., P. Tunestål, and R. Johansson (2008). “Physical modeling and control
of homogeneous charge compression ignition HCCI engines”. In: Proc. Conf.
Decision and Control. Cancun, Mexico, pp. 5615–5620.
Wiebe, I. I. (1970). Brennverlauf und Kreisprozess von Verbrennungsmotoren. VEB
Verlag Technik, Berlin, Germany.
Willems, F., E. Doosje, F. Engels, and X. Seykens (2010). Cylinder Pressure-Based
Control in Heavy-Duty EGR Diesel Engines using a Virtual Heat Release and
Emission Sensor. SAE Technical Paper 2010-01-0564.
Woshni, G. (1967). A Universally Applicable Equation for the Instantaneous Heat
Transfer Coefficient in the Internal Combustion Engine. SAE Technical Paper
670931.
Yang, T., L. Yin, G. Ingesson, P. Tunestål, R. Johansson, and W. Long (2017).
“Simultaneous control of soot emissions and pressure rise rate in gasoline
PPC engine”. In: Proc. IEEE Conf. Control Technology and Applications (CCTA
2017). Hawaii, USA, pp. 572–577.
Yin, L., G. Ingesson, S. Shamun, P. Tunestål, R. Johansson, and B. Johansson
(2015). Sensitivity Analysis of Partially Premixed Combustion (PPC) for Con-
trol Purposes. SAE Technical Paper 2015-01-0884.
Zeldovich, Y. B., P. Y. Sadovnikov, and D. A. Frank-Kamenetskii (1947). Oxidation
of Nitrogen in Combustion. Academy of Sciences of USSR, Institute of Chemi-
cal Physics, Moscow-Leningrad.
Zheng, M., G. T. Reader, D. Wang, J. Zuo, R. Kumar, M. C. Mulenga, U. Asad, D. S.
Ting, and M. Wang (2005). A Thermal Response Analysis on the Transient Per-
formance of Active Diesel Aftertreatment. SAE Technical Paper 2005-01-3885.
Zhou, K. and J. C. Doyle (1998). Essentials of Robust Control. Vol. 104. Prentice
Hall, Upper Saddle River, NJ, USA.

265
Bibliography

Zweigel, R., F. Thelen, D. Abel, and T. Albin (2015). “Iterative learning approach for
diesel combustion control using injection rate shaping”. In: Proc. European
Control Conf. (ECC 2015), pp. 3168–3173.

266
Videos

Videos showing cycle-resolved data for some of the experiments presented are
available in the entry for this thesis in the Lund University Research Portal:
http://portal.research.lu.se/portal/

These videos are also available on the author’s youtube channel:

https://www.youtube.com/channel/UCCsMeCRDrzoJF_wm41X1xuA

Videos of in-cylinder data aid understanding of the controller behavior and

can be used as a complement to the figures presented in the thesis. The videos
present data from the following experiments:

• Closed-loop θ50 and τ experiments in Chapter 7.

• d p max -controller experiments in Chapter 8.

• A comparison between the low-load controllers in Chapter 9.

• Constraint-fulfillment experiments in Chapters 10 and 12.

• Pressure-tracking experiments in Chapter 12.

267