Journal of Process Control 18 (2008) 839–845

Contents lists available at ScienceDirect

Journal of Process Control

journal homepage: www.elsevier.com/locate/jprocont

Coordinating multiple optimization-based controllers: New opportunities

and challenges
James B. Rawlings *, Brett T. Stewart
Department of Chemical and Biological Engineering, University of Wisconsin, Madison, WI, United States

a r t i c l e i n f o a b s t r a c t

Keywords: The status of using many, distributed optimization-based controllers for feedback control of large-scale,
Distributed control dynamic processes is presented and evaluated. We show that modeling the interactions between subsys-
Plantwide control tems and exchanging trajectory information among subsystem model predictive controllers (MPCs) is
Distributed MPC insufficient to provide even closed-loop stability. The cause of this closed-loop instability is competition
Large-scale MPC
between the local agents. We next discuss the cooperative distributed MPC framework, in which the
objective functions of the local MPCs are modified to achieve systemwide control objectives. This
approach provides guaranteed nominal stability and performance properties, but at the cost of a high
degree of communication between the local controllers. We next discuss the issue of taking advantage
of the structure of the connections between the subsystems to reduce the required communication.
The paper concludes by briefly presenting seven current and unsolved research challenges.
Ó 2008 Elsevier Ltd. All rights reserved.

1. Introduction the implementation process uncovers other kinds of flaws that

have escaped detection and the design process is revisited and fur-
Feedback controller design is foremost a design problem and de- ther iterations are required.
sign problems, even engineering design problems, have proven If we examine the history of chemical plant design and opera-
remarkably resistant to precise mathematical formulations with tion in the U.S., we see a steady increase in the complexity in the
unique solutions. Design problems are complex and messy because interactions between the various units comprising the overall
what people want out of a design is complex and messy. Some of plant. This increase was driven by the gain in economic efficiency
the goals for the design are hard to define and subjective, some offered by these more complex and interactive plant design and
of them may be contradictory; at best, the stated goals are usually operation strategies. The chemical plants of the 1950s, which could
incomplete. This does not mean that precise mathematical prob- be operated fully in manual mode by a team of human operators,
lems and tools for computing their solutions are not useful. They would be economically inviable in today’s economy and environ-
are incredibly useful. But their application is iterative. Because ment. As complexity of the plant increased, automatic monitoring
the design goals are complex and difficult to define precisely, the and control systems became a necessity. Multiloop PID control at
design process is almost invariably iterative. The designer or de- the unit level became the first automatic control design. In the last
sign team often proposes a precise, but limited, design problem twenty years, however, centralized MPC of small to medium-sized
statement, solves it exactly and optimally and then inspects the multivariable units has largely replaced multiloop PID in the pro-
outcome, often through numerical evaluation of idealized case cess industries, particularly for the economically important units
studies. Some features are deemed desirable, drawbacks are in the plant.
uncovered, surprises, both good and bad, may be revealed. Inspec- It is interesting to note in passing that the undergraduate con-
tion of the first attempt is meant to inform the designer about the trol curriculum offered in chemical engineering departments has
problem. With this extra knowledge in hand, the designer fashions not kept pace with this most recent transformation in industrial
a better, more comprehensive statement of the design goals and practice. Further discussion of the educational challenges that have
the process continues. At some point the current design is deemed arisen due to this widespread adoption of MPC in control practice
satisfactory and the process of implementation commences. Often are outside the scope of this paper, but control educators are aware
of the problem and are working towards addressing it [1].
We appear now to be at an interesting historical juncture. After
* Corresponding author.
E-mail addresses: [email protected] (J.B. Rawlings), [email protected]
moving quickly from manual control, to multiloop, distributed PID
(B.T. Stewart). control, to centralized, model-based MPC for controlling single and

0959-1524/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2008.06.005
840 J.B. Rawlings, B.T. Stewart / Journal of Process Control 18 (2008) 839–845

small collections of chemical process units, what is the status for need to design a centralized supervisor to accomplish this task. The
improving dynamic performance at the other end of the spectrum: design and maintenance of such a supervisor is no easier than the
the large-scale, integrated collections of many of these units that design and maintenance of the fully centralized controller for
comprise the chemical plant? The complexity and incompletely the large-scale system.
defined nature of the controller design problem again arises. What If we coordinate optimization-based controllers, the local
do we expect of this large-scale controller design? How important agents have a rich structure that we can modify, with predictable
is it that parts of the process can be removed and added back to the and transparent outcomes, to suit overall performance goals. Coor-
overall control system. What model size and complexity is al- dinating paired PID controllers is problematic because the agents
lowed? What is the time scale for online controller decision mak- are so limited. Coordinating MPC controllers starts in a strikingly
ing? How important is it to make an evolutionary transition from different place. The potential is greater. This paper provides an
the current control technology to the next technology. Is a com- overview of some of the opportunities and challenges in coordinat-
plete overhaul and redesign allowed or forbidden? How do the ing MPC controllers.
operations personnel evaluate the various model and controller
maintenance issues that they face? Is a large, centralized control 2. Communication and cooperation among subsystem
system monolithic and difficult to maintain? Or is it easier to main- controllers
tain than collections of smaller models?
In this paper, we do not try to answer all of these difficult ques- We require some terminology and notation to describe the total
tions. We assume instead that completely centralized control is not system, or plant, its decomposition into subsystems, or units, and
likely to be the method of choice for large-scale problems. the components of the controller design problem.
Although we have to let future developments speak to the validity
of this assumption, there is reason to expect a distributed approach 2.1. Models
to large-scale problems to remain a popular choice.1
So in what respect is the current historical juncture interesting? Consider a total system (plant) to be comprised of M intercon-
We have faced the tension between centralized and distributed nected subsystems (units). Let ðyi ðtÞ; ui ðtÞÞ be the ðpi ; mi Þ dimen-
decision making before, in many different contexts. The move from sional vectors of (output, manipulated) variables of the ith
distributed PID to MPC of small systems was essentially a move to- subsystem at time t, in which i 2 IM ¼ ð1; . . . ; MÞ. We assume a fi-
wards centralized decision making. This technology gained support nite dimensional linear time invariant model is suitable to describe
because the performance benefit was large. The main theme in this the dynamics between any manipulated input uj and any measured
paper is that the current situation is interesting because the local output yi
agents are so capable. Imagine trying to coordinate the decision
making of a collection of multiloop PID controllers. If the overall xij ðk þ 1Þ ¼ Aij xij ðkÞ þ Bij uj ðkÞ;
X
system performance is not going well, what are your options to yi ðkÞ ¼ C ij xij ðkÞ; i; j 2 IM
change the behavior? You can modify the local agents by adjusting j

their three knobs: P, I and D. The impact of turning all of these

in which integer k is the sample time. The decentralized model is
knobs up and down on the overall system performance is far from
the set of M models between ui and yi , i 2 IM . The centralized model
obvious. To know, we generally have to do a simulation and a sim-
is the set of M2 models between all inputs and outputs: ui and yj ,
ulation requires starting conditions and gives only a single forecast.
i; j 2 IM . The interaction model for subsystem i is the set of M  1
We would need many simulations to assess the impact of a change.
models between the inputs of other subsystems and outputs of sub-
Choosing which simulations are sufficient to show the entire range
system i: uj and yi j 2 IM ; j 6¼ i. The centralized model is the union of
of behavior is an unsolved problem. Human intervention and judg-
the decentralized model and the M sets of interaction models.
ment during this learning process is way too slow. We can try lar-
ger surgeries: re-configure assignments of measurements with
2.2. Objectives
valves as various types of operational problems are encountered.
A skillful process control engineer who has experience with a spe-
We assume the objective function of each subsystem can be ex-
cific process can do remarkable things with this toolset. But cen-
pressed as a quadratic function of that subsystem’s inputs and out-
tralized decision making based on a multivariable model allows a
puts evaluated at discrete sample times in a prediction horizon
richer set of actions to be evaluated and produces generally better
operation. X
N

Next imagine we wish to coordinate the decision making of a Ui ¼ Li ðyi ðk þ ljkÞ; ui ðk þ ljkÞÞ
l¼1
collection of MPC controllers. How can we modify the behavior
of these local agents? The situation is strikingly different. Each in which L is the stage cost, Li ðy; uÞ ¼ y0 Q i y þ u0 Ri u, Q i ; Ri > 0, and N
agent has a specific model of part of the process that it uses to fore- is the forecast horizon. The forecast of outputs is computed from the
cast outcomes of its decisions. The model is available for the ask- forecast of inputs (the decision variables) under different model
ing. Each agent has a specific cost function that it optimizes to assumptions described below. We assume a suitable objective func-
make its decision. The cost function is available for the asking. Each tion for the total system is a convex combination of the subsystem
agent encodes the differences between its forecast and the mea- objective functions
surement in a specific way. This model is available for the asking. X X
U¼ wi Ui ; wi ¼ 1; wi > 0: ð1Þ
We argue in this paper that the impact of changing the local agents’
i2IM i2IM
models, cost functions, and feedback structures can be reliably as-
sessed without extensive online simulation. Because we know pre-
cisely what the agent is trying to do and what model it is using to 2.3. Communication
do it, we can predictably and reliably alter its behavior. We do not
For the distributed controllers, each control problem is solved
by a local subsystem controller subject to local constraints. We
1
Consider, for example, a centrally planned economy versus a local, market-driven also consider solving these control problems iteratively with a
economy. communication strategy between the iterations of the subsystem
 J.B. Rawlings, B.T. Stewart / Journal of Process Control 18 (2008) 839–845 841

controllers. Because an MPC optimization provides a trajectory of

inputs and not just the current input, we communicate a trajectory
of inputs at each optimization iteration. The iteration of decision
variables is defined as

ðpÞ
upi ¼ wi ui þ ð1  wi Þup1
i ; i 2 IM :

At each iteration, p, the trajectory of inputs is a convex combination Fig. 1. The set U i is the set of upstream subsystems and the set Di is the set of
of the current local optimal solution and the previous iteration. downstream subsystems. Subsystem i is linked to all other subsystems in the
flowsheet by the upstream subsystems l 2 U i and the downstream subsystems
After the iteration converges or the available computation limit is j 2 Di .
reached, the first input in the trajectory is injected into the system
and the next measurement is obtained.

2.4. Controller design

We consider the following four controller design choices [2]. An

agent refers to any optimizer working at the subsystem level.

(1) Centralized control. Single controller. The model is the cen-

tralized model of the total system and the objective function
is the objective function of the total system. Nominal control
performance is optimal.
(2) Decentralized control. M controllers. Each model is the local
subsystem model. Each objective function is the local sub-
system objective function. Design ignores all interactions
between units.
(3) Communication-based control. M controllers. Each model is
the local subsystem model plus interaction models. Each
objective function is the local objective function. For control-
ler i at iterate p, forecast of inputs from controllers j 6¼ i at
iterate p  1 are available.
(4) Cooperative control. M controllers. Each model is the local
unit model plus interaction models. Each objective function
is a copy of the total objective function. For controller i at
iterate p, forecast of inputs from controllers j 6¼ i at iterate
p  1 are available.

3. Good controllers gone bad – densely connected systems

The centralized controller and decentralized controller define

two limiting design extremes. Centralized control accounts for all
possible interactions, large and small, whereas decentralized con-
trol ignores them completely. In decentralized control the local
agents have no knowledge of each others’ actions. It is well known
that the nominal closed-loop system behavior under decentralized
control can be arbitrarily poor (unstable) if the system interactions
are not small. The following reviews provide general discussion of
this and other performance issues involving decentralized control
[3–6].
The next level up in complexity from decentralized control is
communication-based control. In this framework, the agents have
interaction models and communicate at each iteration [7–9]. The
big advantage of communication-based control over decentralized
control is that the agents have accurate knowledge of the effects of
all other agents on their local objectives. The basic issue to analyze
and understand in this setup is the competition between the
agents. That analysis is the subject of noncooperative game theory
[10]. To illustrate some of the issues, consider a simple two-sub-
system problem with scalar inputs and outputs and forecast hori- Fig. 2. Top: Nash equilibrium (n) is stable and near the Pareto optimal solution (p).
Middle: Nash equilibrium is stable but not near the Pareto optimal solution.
zon of N ¼ 1.
Bottom: Nash equilibrium is unstable.
Fig. 2 shows the possible behavior. The Pareto optimal solution
is defined by combining the two local objectives with equal weight.
The Pareto optimal solution is the one found by a centralized con-
troller. The Nash equilibrium is defined as a point satisfying iterating the two local controllers under communication converges
the optimality conditions for each local agent. In the top figure, to the stable Nash equilibrium, which is near the Pareto optimal
842 J.B. Rawlings, B.T. Stewart / Journal of Process Control 18 (2008) 839–845

solution. Communication-based control is likely to provide good The rate of convergence to centralized performance depends
closed-loop system behavior in this scenario. By changing the cost strongly on the type of subsystem interactions – for example, the
functions, we create the middle figure in which the Nash equilib- distillation column requires more iterations than the reactor/sepa-
rium is far from the Pareto solution. The converged solution ob- rator or the unstable system.
tained using a communication-based strategy is far from optimal,
and the closed-loop system using this controller may not be even 3.1. Properties of cooperative MPC
stable. Finally, the bottom figure shows a case in which the Nash
equilibrium is close to the Pareto solution but the Nash equilibrium The following properties of cooperative MPC have been estab-
is not stable. In this case, the communication-based iterates do not lished [2,13].
converge to the Nash equilibrium. The iterates converge to a point
on the boundary of the feasible region, which is far from the Pareto (1) The iterations generated by the cooperative MPC algorithm
solution. Again, closed-loop instability of the communication- are systemwide feasible.
based control system is likely. Note that none of the undesirable (2) Control based on any intermediate termination of the algo-
behavior is caused by a lack of knowledge about the overall system. rithm provides nominal closed-loop stability and zero
All agents have complete information about the effects of all the steady-state offset.
other agents’ actions. (3) If iterated to convergence, the distributed MPC algorithm
Therefore, if the overall system is composed of strongly inter- achieves optimal, centralized MPC control.
acting subsystems, closed-loop instability of decentralized and (4) To handle output instead of state feedback, a distributed
communication-based control seems unavoidable. One alternative estimator design strategy can be implemented, in which
to ensure closed-loop stability, of course, is to use a single central- each estimator is stable and uses only local measurements
ized controller. But there are other alternatives. We can maintain to estimate subsystem states. The combined distributed esti-
the distributed structure of the M local controllers, but change mator-distributed regulator is feasible and closed-loop sta-
the objective functions so that the local agents cooperate. Changing ble for all iteration numbers in the case of decaying
the cost function is a simple matter of rewriting the data in the lo- estimate error.
cal agents QP subproblems. In fact, the data for various levels of
cooperation, varying from decentralized control to fully coopera-
tive control, can be stored so that different control scenarios can 4. Topology of typical chemical processes
be loaded and implemented. This approach provides an evolution-
ary path from a current decentralized technology to something In Section 2 an MPC cooperation strategy with guaranteed per-
approaching centralized control, but without removing the local formance was summarized. Yet, while ensuring stability and cen-
control structure that may already be in place. tralized-like behavior, it requires a completely connected
Between these alternatives lies coordinator MPC [11]. This communication strategy. Every agent in the plant communicates
semi-centralized control strategy partitions the decision variables with all the others. This level of communication makes sense be-
between local and centralized controllers. Economically important cause each subsystem in the plant, in open loop, may affect all
inputs are optimized plant-wide; others are left to local control. the others. This communication is not desirable, however, because
This solution is applicable if the plant is economically optimal at it requires the final unit to communicate with the initial unit, even
maximum throughput. It is inflexible to bottleneck changes, how- if these processes are connected only through many intermediate
ever, and the plant-wide calculation adds complexity to the control units. This drawback motivates developing a strategy in which
topology. Cooperative MPC maintains the distributed control struc- closed-loop stability, at least, is guaranteed but unnecessary com-
ture common to chemical plants. munication is eliminated. In this section we focus on the character-
Venkat [12] studied a number of small typical chemical process ization of such a strategy.
examples and found it is rather easy to generate closed-loop insta- In a typical chemical process, subsystems are connected
bility for decentralized and communication-based control systems. through material, energy, and information flows. These flows gen-
Table 1 summarizes some of these results. The examples are: (i) a erally pass from subsystem to subsystem, so that each subsystem
2  2 transfer function of a distillation column, (ii) a reactor/sepa- directly interacts only with its nearest neighbors. Interactions be-
rator obtained by linearizing a nonlinear, fundamental model at a yond nearest neighbors occur through intermediate subsystems.
desired steady state, and (iii) a an open-loop unstable mathemati- Therefore, non-nearest neighbor interactions may be modeled as
cal example. Decentralized and communication-based MPC are the product of multiple nearest neighbor interactions in series.
unstable for the first two examples and increase the cost compared Also, in many sections of the plant, a given subsystem directly
to centralized MPC by almost 100% in the third example. Coopera- interacts only with its downstream subsystems. For example, in
tive MPC is always closed-loop stable. The nominal performance of Fig. 3, subsystem 1 affects subsystem 2 directly, but affects subsys-
the closed-loop system improves with the number of iterations of tem 4 by way of subsystems 2 and 3. This topology may be
the local controllers and converges to the centralized solution. exploited to reduce the amount of communication required for
cooperative control.

Table 1
Closed-loop performance comparison of centralized MPC (Cent-MPC), decentralized
MPC (Decent-MPC), communication-based MPC (Comm-MPC) and cooperative MPC
(Coop-MPC)

Dist. col. React/sep. OL unstable

Kcost DKcost % Kcost DKcost % Kcost DKcost %
Cent-MPC 1.72 0 2.0 0 1.78 0
Decent-MPC 1 1 1 1 3.53 98.3
Comm-MPC 1 1 1 1 3.53 98.2
Coop-MPC (1 iter.) 6.35 269.2 2.13 6 2.03 13.9 Fig. 3. Ethylene glycol flowsheet. (1) Feed tank; (2) preheater; (3) reactor (4)
Coop-MPC (10 iter.) 1.74 1.32 2.01 0.09 1.8 0.8 evaporators; (5) light end columns; (6) mono ethylene glycol column; (7) higher
glycol recovery.
 J.B. Rawlings, B.T. Stewart / Journal of Process Control 18 (2008) 839–845 843

Given the structure above, a different interaction model can be yields an important effect: the synchronization step can be solved
derived. This model assumes the states of a given subsystem are a locally on each subsystem. To synchronize, subsystem i must know
function of only the subsystem’s states and inputs and the states of only the states, already available from the optimization step, of the
the upstream subsystems as in Fig. 1. Specifically subsystems upstream. Therefore no extra communication is
X needed. The reduced communication requirement in distributed
xi ðk þ 1Þ ¼ Aii xi ðkÞ þ Ail xl ðkÞ þ Bi ui ðkÞ ð2Þ MPC in this case is analogous to the relative gain array (RGA)
l2U i becoming a diagonal matrix for a triangular matrix of transfer
functions [14, p. 737].
in which U i  IM is the set of nearest neighbor subsystems upstream A typical plant has many recycle streams, and the recycle
of subsystem i. This model implies a fundamentally different set of streams may be difficult to identify, as in the ethylene glycol exam-
interactions than used in Section 2 in which input-to-state interac- ple in Fig. 3. How can communication be reduced in this situation?
tions are considered. The state-to-state model (2) is equivalent We propose a hybrid strategy. Total communication, as in Section
however. The equivalence is revealed by recursively substituting 2, is implemented between subsystems involved in the recycle
for all xl ðkÞ into xi ðk þ 1Þ, and recovering a strict input-to-state mod- while reduced communication, described in this section, is used
el. More importantly, by considering this state-to-state model, com- between subsystems not in the recycle. The total communication
munication is reduced. Defining the set U i also defines the set of and reduced communication areas exchange information and iter-
subsystems that directly affect subsystem i. Typically U i is a subset ate in parallel. For example, in Fig. 4, subsystems 3–5 are involved
of the entire plant’s subsystems, which reduces the interactions that in recycle, so full communication is used between these subsys-
must be considered. tems. Between all others, reduced communication may be used.
In MPC, each optimizer is given the task of predicting the effect Similar techniques are used in the optimal matching and shrinking
of its inputs on a given objective. In cooperative MPC, this objective algorithms in graph reduction [15, pp. 127–198] and in the analy-
is the performance of the entire plant U. As in Section 2 the entire sis of electrical networks [16, pp. 263–280]. The analysis of this
plant’s objective is the convex sum of the subsystems’ objective Ui . simple example is straightforward, but more complicated flow-
Using Eq. (2) to predict how the inputs of subsystem i affect the sheet topologies may require general and sophisticated tools. In
plant objective and considering nearest neighbor interactions only, Fig. 3, for example, must subsystem 1 communicate with subsys-
the summation in (1) is over i and the subsystems downstream tem 4 directly, or is it sufficient to communicate only with subsys-
from i tem 2?
X Consider the extremes of the recycle problem. Fig. 4 represents
/i ¼ wr Ur ð3Þ
the most basic recycle, in which flow moves, on average, down-
r2fi;Di g
stream with a recycle occurring occasionally. This periodic recycle
in which /i is the reduced centralized objective for subsystem i and may be controlled via the hybrid strategy outlined above. Fig. 5 is
Di is the set of nearest neighbor subsystems downstream from sub- the other extreme of recycle. Subsystem 1 is affected directly by
system i. The reduction of terms in the summation follows from subsystem 6, the final unit in the flowsheet. Must all subsystems
examining each term. If no decision variables appear in a term of communicate in this example or is there a more elegant way to re-
the summation, then this term is constant and does not affect the duce communication? In general, a plant is a combination of these
solution. Using (2) for the prediction of the plant objective, the deci- recycle extremes. The next step in this line of research is to derive a
sion variable ui ðkÞ appears in (1) only in the ith term and jth terms framework for identifying these recycles and specifying the neces-
such that i 2 U j . The latter may be restated as j 2 Di . After substitut- sary communication strategies to handle them.
ing (2) into (3) the MPC subproblem is obtained. The controller for
subsystem i must consider the state trajectories of the upstream 5. Future challenges
subsystems U i and the input trajectories of the downstream subsys-
tems Di . Qualitatively, in order to make an optimal input decision, The following issues represent open research challenges. Pro-
subsystem i must know where the upstream subsystems are going gress on any of these issues will likely further the development
and be able to forecast the downstream subsystems. But notice it of a more comprehensive and reliable controller design strategy
does not need a forecast of any other subsystem. This behavior is suitable for the large-scale, and challenging applications faced by
acceptable because the other subsystems make these forecasts, practitioners.
and account for these non-nearest neighbor interactions.
This reduction of communication does not come for free, how- 5.1. Exploiting the structures in application domains
ever. The MPC optimization returns an optimal trajectory of inputs.
The states, which are communicated after local optimizations, As a research community, we have just begun to think about the
must be synchronized with the optimal inputs to obtain the opti- types of structure arising in applications. The topological connec-
mal state trajectory. In general, these states cannot be obtained tions of chemical processes described in Section 4 is one such
using the local model. The trouble arises in the presence of recycle. example. Recognizing and exploiting these kinds of structures
Consider the ethylene glycol plant in Fig. 3 in which subsystems 1 may prove critical to success in many fields.
and 2 are involved in a recycle. According to Eq. (2), the states of
subsystem 1 affect subsystem 2, but at the same time the states
of subsystem 2 affect subsystem 1. These states must then be ob-
tained simultaneously. Therefore an extra step, and extra time, is
needed to iteratively synchronize the states, or one of the MPCs
solves for the states of both systems. Adding time to the optimiza-
tion is never a good idea, however, and, in a distributed optimiza-
tion, allowing one of the subsystems to solve for all states is not
much different than centralized control.
Consider the special case without recycle. In this case, material
flows only downstream and the problem of downstream subsys- Fig. 4. Periodic recycle. Flow is mostly in the direction from the first subsystem to
tems affecting upstream subsystems is avoided. This property the last with local recycle occurring periodically throughout the flowsheet.
844 J.B. Rawlings, B.T. Stewart / Journal of Process Control 18 (2008) 839–845

interesting. Construction of disturbance invariant sets [21,22] for

each subsystem could prove useful to establish robust stability.

5.6. Partial cooperation

Partial cooperation is a method that allows the control designer

to enforce limitations on the ways optimization-based controllers
use their available inputs to meet performance specifications
[23]. It has the nice side benefit of reducing communication among
subsystem controllers. But closed-loop properties for the partial
cooperation framework are not available. Techniques for further
reducing communication among subsystems, without compromis-
ing closed-loop stability, should be investigated.

5.7. Time-scale separation and fast subsystems

Fig. 5. Total recycle. The last subsystem in the process flows into first.

To implement cooperative distributed MPC for systems with

5.2. Coupled input constraints fast sampling rates, one may require techniques that allow a quick
evaluation of the MPC optimization problem. The possibility of
The FC-MPC algorithm converges to the optimal centralized solu- employing explicit MPC techniques [24–27] for distributed MPC
tion when the input constraints are uncoupled between subsys- should be investigated. One complication in distributed MPC is
tems. For input constraints coupled between subsystems, nominal that the input trajectories for interconnected subsystems’ MPCs
closed-loop stability is still guaranteed, but performance remains are additional parameters for each MPC optimization problem.
an open question. At present we do not have any bounds on the de- The dimensionality of the parameter space consequently, is much
gree of suboptimality that can be caused by coupled input greater in distributed MPC.
constraints. Zero-order holds have been the method of choice in centralized
MPC applications. First-order holds offer many advantages for fast
5.3. Communication disruption systems. For the same control performance, a first order hold
allows a larger sampling time than a zero order hold; this feature
The goal of a distributed control design system is to maintain a by itself allows further iteration of the distributed MPC system.
high level of overall system performance with as little communica- At this time we know little about the advantages and disadvan-
tion as possible between the subsystems. Reliable strategies are re- tages of first-order holds in asynchronous, distributed MPC.
quired for handling possible disruptions and delays in the
communication of input trajectories among subsystems. It is ex-
pected that the closed-loop system can be destabilized with incor- 6. Conclusion
rect input trajectory information due to information loss or delays.
It would be desirable to have lower performance backup strategies Given the long introduction, the conclusion can be brief. We
available if communication fails completely. A plant’s previous have argued that coordinating optimization-based controllers
decentralized control system may be suitable for this task, but offers interesting capabilities for tackling control system design
other options may be investigated such as those considered in for large-scale processes. Optimization-based local agents operate
recent work on control over networks [17–19]. at a high level. By modifying their objective functions, the agents
cooperate. By modifying their interaction models, the agents are
5.4. Closed-loop identification of subsystem interactions aware of the impact of their decisions on other parts of the overall
system. By taking advantage of the process topology, we can re-
The distributed MPC framework requires both subsystem mod- duce the required communication between the agents. There is
els and models of the important interactions between subsystems. no free lunch. If the system is highly interactive, it requires a high
While closed-loop identification is a well-studied field, tailoring degree of communication among the local agents, or a centralized
the existing techniques for distributed MPC is a relatively recent controller. But a highly interactive overall system is unlikely to be
research area [20]. Improvements in techniques for closed-loop the rule in large-scale chemical processes. Chemical process design
identification for distributed MPC will likely prove critical for prac- does not typically produce this kind of highly interactive overall
tical implementation. Reliable integration of the algorithm used for system. Material, energy and information generally flows sequen-
closed-loop identification with the algorithm for distributed MPC tially from unit to unit with smaller levels of recycle and integra-
may have significant impact in the process industries. tion providing the major interactive coupling. Because controller
design for large-scale systems is a complex design problem, many
5.5. Robustness to model errors interesting research issues remain. These include: exploiting sys-
tem structure, handling input constraint coupling and disruption
Handling uncertainty in the controller model remains a key is- in communication, identifying interaction models online, ensuring
sue that needs to be addressed. Interaction models are typically robustness to model errors, imposing controller restrictions, and
identified using closed-loop operating data. Typically, for small treating time-scale separation.
plant-model mismatch, the feedback in standard MPC is adequate Further research in this field will benefit greatly from close con-
to obtain good closed-loop performance. When the plant-model tact between theory and practice. The major downside risk to avoid
mismatch is more significant, robust distributed MPC design may is producing elegant controller design methodologies that no one
be necessary. With this issue in mind, a thorough investigation into uses. Because this line of research is motivated by practical needs,
robustness theory for distributed MPC needs to be undertaken. close collaboration between researchers and practitioners is a pre-
Establishing properties such as robust feasibility and stability in requisite for evaluating progress and defining new research
the distributed MPC setting could prove to be both useful and opportunities.
 J.B. Rawlings, B.T. Stewart / Journal of Process Control 18 (2008) 839–845 845

Acknowledgments [13] A.N. Venkat, J.B. Rawlings, S.J. Wright, Stability and optimality of distributed,
linear MPC. part 2: Output feedback, Technical Report 2006–04, TWMCC,
Department of Chemical and Biological Engineering, University of Wisconsin–
The authors would like to thank Dr. A.N. Venkat and Professors Madison, October 2006b. <http://jbrwww.che.wisc.edu/tech-reports.html>.
S.J. Wright and I.A. Hiskens for helpful discussion of these ideas. [14] B.A. Ogunnaike, W.H. Ray, Process Dynamics, Modeling, and Control, Oxford
University Press, New York, 1994.
The authors gratefully acknowledge the financial support of the
[15] W.J. Cook, W.H. Cunningham, W.R. Pulleyblank, A. Schrijver, Combinatorial
NSF through Grant #CTS-0456694, and the industrial members of Optimization, Wiley-Interscience Series in Discrete Mathematics and
the Texas–Wisconsin Modeling and Control Consortium. Optimization, Wiley-Interscience, New York, 1998.
[16] S. Chan, Introductory Topological Analysis of Electrical Networks, Holt,
Rinehart and Winston Series in Electrical Engineering, Electronics, and
References Systems, Holt, Rinehart and Winston, Inc., New York, 1969.
[17] G. Baliga, P.R. Kumar, A middleware for control over networks, in: Proceedings
[1] T.F. Edgar, B.A. Ogunnaike, J.J. Downs, K.R. Muske, B.W. Bequette, Renovating of the Joint 44th IEEE Conference on Decision and Control and European
the undergraduate process control course, Comput. Chem. Eng. 30 (2006) Control Conference, Seville, Spain, 2005, pp. 482–487.
1749–1762. [18] A. Casavola, M. Papini, G. Franzé, Supervision of networked dynamical systems
[2] A.N. Venkat, J.B. Rawlings, S.J. Wright, Stability and optimality of distributed, under coordination constraints, IEEE Trans. Autom. Contr. 51 (3) (2006) 421–
linear MPC. part 1: State feedback, Technical Report 2006–03, TWMCC, 437.
Department of Chemical and Biological Engineering, University of Wisconsin– [19] O.C. Imer, S. Yuksel, T. Basar, Optimal control of dynamical systems over
Madison, October 2006a. <http://jbrwww.che..wisc.edu/tech-reports.html>. unreliable communication links, in: IFAC Symposium on Nonlinear Control
[3] D.D. Siljak, Decentralized Control of Complex Systems, Academic Press, Systems, Stuttgart, Germany, 2004, pp. 1521–1524.
London, 1991. [20] R.D. Gudi, J.B. Rawlings, Identification for decentralized model predictive
[4] J. Lunze, Feedback Control of Large Scale Systems, Prentice-Hall, London, UK, control, AIChE J. 52 (6) (2006) 2198–2210.
1992. [21] I. Kolmanovsky, E.G. Gilbert, Theory and computation of disturbance invariant
[5] T. Larsson, S. Skogestad, Plantwide control – a review and a new design sets for discrete-time linear systems, Math. Prob. Eng. 4 (4) (1998) 317–367.
procedure, Mod. Ident. Contr. 21 (4) (2000) 209–240. [22] S.V. Rakovic, E.C. Kerrigan, K.I. Kouramas, D.Q. Mayne, Invariant
[6] H. Cui, E.W. Jacobsen, Performance limitations in decentralized control, J. Proc. approximations of robustly positively invariant sets for constrained linear
Contr. 12 (2002) 485–494. discrete-time systems subject to bounded disturbances, Tech. Rep.,
[7] D. Jia, B.H. Krogh, Min–max feedback model predictive control for distributed Department of Engineering, University of Cambridge, Cambridge, UK, cUED/
control with communication, in: Proceedings of the American Control F-INFENG/TR.473 (January 2004).
Conference, Anchorage, Alaska, 2002, pp. 4507–4512. [23] A.N. Venkat, J.B. Rawlings, S.J. Wright, Distributed model predictive control of
[8] N. Motee, B. Sayyar-Rodsari, Optimal partitioning in distributed model large-scale systems, in: Assessment and Future Directions of Nonlinear Model
predictive control, in: Proceedings of the American Control Conference, Predictive Control, Springer, 2007, pp. 591–605.
Denver, Colorado, 2003, pp. 5300–5305. [24] A. Bemporad, C. Filippi, Suboptimal explicit receding horizon control via
[9] W.B. Dunbar, R.M. Murray, Distributed receding horizon control with approximate multiparametric quadratic programming, IEEE Trans. Autom.
application to multi-vehicle formation stabilization, Automatica 42 (4) Contr. 117 (1) (2003) 9–38.
(2006) 549–558. [25] A. Bemporad, M. Morari, V. Dua, E.N. Pistikopoulos, The explicit linear
[10] T. Basßar, G.J. Olsder, Dynamic Noncooperative Game Theory, SIAM, quadratic regulator for constrained systems, Automatica 38 (1) (2002) 3–20.
Philadelphia, 1999. [26] G. Pannocchia, J.B. Rawlings, S.J. Wright, Fast, large-scale model predictive
[11] E.M.B. Aske, S. Strand, S. Skogestad, Coordinator MPC for maximizing plant control by partial enumeration, Automatica 43 (2007) 852–860. <http://
throughput, Comput. Chem. Eng. 32 (2008) 195–204. www.elsevier.com/locate/automatica>.
[12] A.N. Venkat, Distributed model predictive control: theory and applications, [27] P. Tondel, T.A. Johansen, A. Bemporad, An algorithm for multi-parametric
Ph.D. thesis, University of Wisconsin–Madison, October 2006. <http:// quadratic programming and explicit MPC solutions, Automatica 39 (3) (2003)
jbrwww.che.wisc.edu/theses/venkat.pdf>. 489–497.