FEATURE SERIES: PRACTICAL PROCESS CONTROL

FEATURE SERIES PRACTICAL PROCESS CONTROL

2: Determining
Process Dynamics
Myke King continues his detailed series on process control, seeking to inspire
chemical engineers to exploit untapped opportunities for improvement

I
DENTIFYING process dynamics is an essential first step in this case thermal capacity. It is responsible for the second of
to achieving good control design. Many control engi- our dynamic parameters which we describe as lag (τ). The water
neers will be guilty of spending many hours tuning temperature has changed from cold to hot, from one steady state
controllers by trial and error. This can largely be to another. We describe this difference as the steady state change.
avoided. Using the process dynamics, obtained from
straightforward plant testing, simple calculations can then be
applied to identify optimum tuning for all types of controller. PROCESS GAIN
So, imagine that you’re the first into the bathroom in the Now we must leave the analogy because we can’t use the
morning and you turn on the hot tap. Initially it produces cold opening of the tap to actually control the temperature. We’ll
water and does so for some time. We anticipate this behav- consider instead the fired heater shown in Figure 1. It has a flow
iour; the water, held up between the source of hot water and controller (FC) installed on the fuel supply and temperature
the tap, has had time to cool to room temperature. This water indicator (TI) on the product outlet. Our intent is to install a
first has to be displaced. It has to be transported to the tap temperature controller that will adjust the fuel flow. To design
and causes what we describe as a transport delay. Also known as this controller we first determine the process dynamics. We
deadtime ( ), it is the first of the dynamic constants that we use make a step change to the FC set-point. The temperature shows
to describe process behaviour. We also notice that, when the much the same response as our bathroom tap. The TI is located
cold water has been displaced, the water temperature doesn’t downstream of the firebox and so exhibits transport delay. The
immediately increase to that of the hot water; it ramps up to heating coil, and the fluid it contains, have significant capacity
it. Part of the energy in the water is required to warm up the to absorb heat; so we also see lag. The final, and most impor-
pipework leading to the tap. The process is said to have capacity, tant parameter, is the process gain (Kp). This is defined as the

Figure 1: Process response

70

60
PROCESS PARAMETER (% OF RANGE)

50

40
FUEL FLOW ∆PV
SETPOINT
∆MV
TI 30

20

TEMPERATURE
FC 10

0
0 5 10 15 20 25

TIME (MINUTES)

MARCH 2023 | The Chemical Engineer | PAGE 40

Process Control 981 AD.indd 40 23/02/2023 17:42:24

 FEATURE SERIES PRACTICAL PROCESS CONTROL

steady state change in temperature, our process variable (PV), manipulating the flow of cooling water, the process gain would
divided by the steady state change in fuel flow set-point, our be negative. The choice of instrument ranges will often result
manipulated variable (MV). in Kp being close to 1 but there can be occasions where it differs
by several orders of magnitude.

We need to be careful with units. The process gain, as we’ll ESTIMATING THE TIME CONSTANTS
see in a future article on tuning, is largely used to determine There are numerous published methods for determining and
the controller gain (Kc). We’ll see that controllers installed in τ. Most are based on the assumption that the response curve
programmable logic controllers (PLC) or distributed control can be represented by
systems (DCS) work in dimensionless form. Kc must be dimen-
sionless and so, therefore, must Kp. To make both PV and MV
dimensionless; we divide each by its corresponding range (or To illustrate the principle behind some of these methods, we set
span). The instrument ranges for the TI and FC will have been the time elapsed since expiry of the deadtime (t) to τ. This gives
assigned by the instrument engineer as part of the control
system configuration. They usually remain constant. (The
need to retune the controller, if the instrument ranges are ever We can use this formula to determine ; it is the time taken
changed, is frequently overlooked.) (after the deadtime has elapsed) for the PV to reach 63% of the
steady state change. Knowing only + , we of course need to
also determine . One approach is to identify two points on the
response curve. One of the more reliable methods is based on
identifying t 25 and t75– the times taken to reach 25% and 75% of
the steady state change. Fitting our exponential curve through
Note that control applications, like multivariable predictive these two points gives
controllers (MPC), residing at the computer level generally
operate in engineering units. For these applications, Kp should
not be converted to its dimensionless form. Numerically Kp
can be positive or negative, even zero in very unusual circum-
stances. In our example of the fired heater, increasing fuel Numerically, can be negligibly small or many hours. However,
increases the temperature and so the process gain is positive. with digital control, it will never be truly zero because of
If we were planning to control the temperature of a cooler by the additional delay caused by the controller scan interval.

Using the
process dynamics,
obtained from
straightforward
plant testing, simple
calculations can then
be applied to identify
optimum tuning for
all types of controller

MARCH 2023 | The Chemical Engineer | PAGE 41

Process Control 981 AD.indd 41 23/02/2023 17:42:31

 FEATURE SERIES PRACTICAL PROCESS CONTROL

Similarly τ can vary greatly from process to process. The unit this we need to collect data at more than the two steady state
of measure chosen should be either minutes or seconds. The conditions. So, if we first increased the fuel set-point, we
choice will depend on the units used by the control system for should return to the starting condition and then test with an
the controller tuning parameters. equivalent decrease – using this to give a second estimate of Kp.
No process is truly linear. The design technique we will cover
in the article on controller tuning gives a robust controller that
ORDER will tolerate variation in Kp of ±20%. If, from our plant testing,
However, a process response will usually be more complex than the higher estimate of Kp is less than 1.5 times the lower, then
that described by a first order exponential function. There will both estimates will be within 20% of the average and we can
be multiple sources of lag. For example, a small lag (τ1) will treat the process as linear. If not, then we will need to apply
be introduced by the flow controller; although we change its one of the methods we cover in the forthcoming articles on
set-point as a step, the valve is actually moved by the control- signal conditioning.
ler. The firebox will introduce a large lag (τ2) by absorbing some
of the energy provided by the additional fuel. The coil will
behave similarly (τ3). The thermocouple is located in a ther- CURVE FITTING
mowell, which again has (albeit small) heat capacity that lags The limitation of most published methods is that they are
the temperature measurement (τ4). While not precisely correct, applicable to a single step change, with the process starting
we can think of the dynamic behaviour as governed by and finishing at steady state. In practice, effective step-test-
ing comprises a series of step tests of varying size and duration
– covering the normal operating range. This provides a more
reliable result. Further, a series of small steps is likely to be more
Figure 2 shows the cumulative effect of what is now a fourth acceptable, to the process operator, than a single large one. It is
order process. It is, however, unrealistic that we can quantify advisable to wait for steady state for at least some of the steps. But
individual lags. Instead we take the lumped parameter approach, we might, particularly on processes with large time constants,
making the assumption that the process is first order. Included reduce testing time by taking the next step without waiting. But
in Figure 2 is the result of applying the two-point method to the main issue is that we don’t want to be constrained to making
model identification. This approximation is usually reliable step changes. We’ll see later that we will require the dynamics of
enough to design the controller. There are, however, a few other variables that might disturb the process. It may be impos-
exceptions that we will cover in future articles. sible to change these as steps, most notably if we want to include
ambient temperature. But the most compelling argument is that
most controller tuning is done on existing controllers. In our case
LINEARIT Y the heater outlet temperature controller may have been in service
The underlying assumption, so far, is that the process is linear, for some time, but we see the potential to improve its perfor-
i.e. there is a linear relationship between PV and MV– meaning mance. It’s preferable to perform step-testing with the controller
that Kp (the slope of this relationship) is constant. To confirm on auto or in closed loop mode. We maintain control of the process
during testing and we can stay within pre-defined bounds for
the PV. One might think that the temperature controller tuning
will somehow affect the model identification. By changing its
Figure 2: First order approximation
set-point we cause it to change the fuel flow set-point. Certainly,
the way it adjusts the flow set-point will depend on the tempera-
1.2 ture controller tuning and the changes will certainly not be
FLOW SET-POINT
step changes. But this does not affect the relationship between
1.0 temperature and fuel flow. It does, however, preclude the use
FRACTION OF STEADY STATE

FUEL FLOW
of any identification technique which assumes a step change.
0.8
COIL TEMPERATURE
Instead we must apply some form of curve fitting.
At this stage we move from analog to digital. Controllers are
0.6 FLUID TEMPERATURE now largely digital and, although the process is analog, data
collected are stored digitally and at a fixed interval. The digital
0.4 MEASURED TEMPERATURE
equivalent of a first order process predicts behaviour as

0.2
FIRST ORDER APPROXIMATION
0
TIME The current PV*n is predicted from the previous value (PV*n-1) and
the delayed , where ts is the data collection interval.

MARCH 2023 | The Chemical Engineer | PAGE 42

Process Control 981 AD.indd 42 23/02/2023 17:42:32

 FEATURE SERIES PRACTICAL PROCESS CONTROL

The bias is required because it is unlikely that the PV will be using the formulae above. The sum of the squares of the predic-
zero when the MV is zero. Because this is a sampled system, tion error is minimised by adjusting Kp, , τ and bias. Plotting the
is restricted to integer values. To overcome this limitation predicted PV against the measured PV will highlight any nonline-
we define the delayed MV as an interpolation between adjacent arity. Commercial model identification software is widely available
values. but the calculations above can readily be implemented in Excel,
using its Solver function. An example spreadsheet may be down-
loaded at www.whitehouse-consulting.com/icheme/model.xlsx.
One of the advantages of this approach is that it can be applied
Figure 3 shows the calculation methodology. From the process to set-point changes made routinely in the past. As a routine,
data historian, MV and PV data are collected along with their time the data historian should be configured to include the set-point,
stamp (in this case in seconds). PV*0 is initialized to the actual measurement and output of all controllers. Data compression, used
measurement (PV0). The remaining predictions are determined by many historians, should be disabled since this will distort the
process dynamics.

Figure 3: Calculation procedure

NEXT ISSUE
TIME MV PV* PV The next article will be one of several on the subject of the
0 MV0 PV0 PV0
PID control algorithm. We’ll describe the many modifications
ts MV1 • PV1 (PV1-PV*1)2
made to the classic version and how they are configured in
2ts MV2 • PV2 (PV2-PV*2)2
the leading control systems.
3ts MV3 • PV3 (PV3-PV*3)2
• • • • •
• MVn- /ts • • •
EQUATION
• • OF FIRST • • • Myke King CEng FIChemE is director of Whitehouse Consulting, an
• • ORDER PV*n-1 • •
MODEL independent advisor covering all aspects of process control. The
n.ts MVn PV*n PVn (PVn-PV*n)2 topics featured in this series are covered in greater detail in his book
• • • • •
Process Control – A Practical Approach, published by Wiley in 2016.
• • ADJUST • • •
K p, ,τ
AND bias ∑(PVi-PV*i)2 Disclaimer: This article is provided for guidance alone. Expert
engineering advice should be sought before application.

IChemE Global Awards 2023

Celebrating Chemical Engineering Excellence

Open for entries

With 19 categories, our Global Awards celebrate the best people, projects
and companies from across the process industries, worldwide.
Entry is free and open to IChemE members and non-members.
Winners will be announced at the Hilton Birmingham Metropole, UK,
on 30 November 2023.

www.icheme.org/globalawards
TCE957_21

TCE MAR Awards HPSM AD.indd 1 23/02/2023 14:33:33

MARCH 2023 | The Chemical Engineer | PAGE 43