Professional Documents
Culture Documents
PDC7
PDC7
7. PID Controllers
Control
Laboratory
7.0 Overview
7.1 PID controller variants
7.2 Choice of controller type
7.3 Specifications and performance criteria
7.4 Controller tuning based on frequency response
7.5 Controller tuning based on step response
7.6 Model-based controller tuning
7.7 Controller design by direct synthesis
7.8 Internal model control
7.9 Model simplification
Process
Control
7.0 Overview
Laboratory
PID controller (”pee-i-dee”) is a generic name for a controller containing a linear
combination of
proportional (P)
integral (I)
derivative (D)
terms acting on a control error (or sometimes the process output).
All parts need not be present. Frequently I and/or D action is missing, giving a
controller like
P, PI, or PD controller
It has been estimated that of all controllers in the world
95 % are PID controllers
Process
Control
7.1 PID controller variants
Laboratory
7.1.1 Ideal PID controller
An ideal PID controller is described by the control law
1 𝑡𝑡 d𝑒𝑒(𝑡𝑡)
𝑢𝑢 𝑡𝑡 = 𝐾𝐾c 𝑒𝑒 𝑡𝑡 + ∫ 𝑒𝑒 𝜏𝜏 d𝜏𝜏 + 𝑇𝑇d + 𝑢𝑢0 (7.1)
𝑇𝑇i 0 d𝑡𝑡
𝑢𝑢(𝑡𝑡) is the controller output
𝑒𝑒 𝑡𝑡 = 𝑟𝑟 𝑡𝑡 − 𝑦𝑦(𝑡𝑡) is the control error, which is the difference between the
setpoint 𝑟𝑟(𝑡𝑡) and the measured process output 𝑦𝑦(𝑡𝑡)
𝐾𝐾c is the proportional gain
𝑇𝑇i is the integral time
𝑇𝑇d is the derivative time
𝑢𝑢0 is the “normal” value of the controller output
The transfer function of the PID controller is
𝑈𝑈(𝑠𝑠) 1 𝐾𝐾c
𝐺𝐺PID = = 𝐾𝐾c 1 + + 𝑇𝑇d 𝑠𝑠 = 1 + 𝑇𝑇i 𝑠𝑠 + 𝑇𝑇i 𝑇𝑇d 𝑠𝑠 2 (7.2)
𝐸𝐸(𝑠𝑠) 𝑇𝑇i 𝑠𝑠 𝑇𝑇i 𝑠𝑠
Depending on the values of 𝑇𝑇i and 𝑇𝑇d , the transfer function of the PID controller
Process
Control can have
Laboratory
real or complex-valued zeros
Complex zeros might be useful for control of underdamped systems with complex
poles.
Process
7.1.2 The series form of a PID controller
Control
Process
7.1.3 A PID controller with derivative filter
Control
A drawback with the ideal PID controller (7.1) is that the derivative part cannot
Laboratory
be realized exactly in a real controller. For example, if the control error 𝑒𝑒(𝑡𝑡)
changes as a step, the derivate in (7.1) becomes infinitely large. This problem can
be remedied by
filtering the signal to be differentiated.
This also has the practical advantage that (high-frequency) noise is filtered before
differentiation.
The transfer function of a parallel form PID controller with a derivative filter is
1 𝑇𝑇d 𝑠𝑠
𝐺𝐺PIDf = 𝐾𝐾c 1 + + (7.6)
𝑇𝑇i 𝑠𝑠 𝑇𝑇f 𝑠𝑠+1
The transfer function of a series form PID controller with a derivative filter is
usually stated in the form
1 𝑇𝑇d′ 𝑠𝑠+1
𝐺𝐺PIPDf = 𝐾𝐾c′ 1+ (7.7)
𝑇𝑇i′ 𝑠𝑠 𝑇𝑇f′ 𝑠𝑠+1
𝑇𝑇f and 𝑇𝑇f′ are filter constants, usually 10-30 % of corresponding derivative time.
For calculation of the parameters of the series form from the parameters of the
parallel form, we define the parameter
4𝑇𝑇i (𝑇𝑇d +𝑇𝑇f )
𝛿𝛿 = 1 − (7.9)
(𝑇𝑇i +𝑇𝑇f )2
If 𝛿𝛿 ≥ 0, the zeros of the parallel PID are real. Then, there exists a series-form PID
controller which is equivalent to the parallel form according to
(𝑇𝑇i +𝑇𝑇f ) 𝑇𝑇i′
𝑇𝑇i′ = 1 + 𝛿𝛿 , 𝑇𝑇d′ = 𝑇𝑇i + 𝑇𝑇f − 𝑇𝑇i′ , 𝑇𝑇f′ = 𝑇𝑇f , 𝐾𝐾c′ = 𝐾𝐾c (7.10)
2 𝑇𝑇i
Process
7.1.4 Differentiation of the measured output
Control
Even if we have a derivative filter, a step change in the setpoint 𝑟𝑟(𝑡𝑡) tends to
Laboratory
affect the derivative part much more strongly than a disturbance in the output
𝑦𝑦(𝑡𝑡). A remedy to this is to
differentiate the (filtered) output instead of the control error 𝑒𝑒(𝑡𝑡).
The ideal control law (7.1) then becomes
1 𝑡𝑡 d𝑦𝑦f (𝑡𝑡)
𝑢𝑢 𝑡𝑡 = 𝐾𝐾c 𝑒𝑒 𝑡𝑡 + ∫ 𝑒𝑒 𝜏𝜏 d𝜏𝜏 − 𝑇𝑇d + 𝑢𝑢0 (7.12a)
𝑇𝑇i 0 d𝑡𝑡
d𝑦𝑦f (𝑡𝑡)
𝑇𝑇f + 𝑦𝑦f 𝑡𝑡 = 𝑦𝑦(𝑡𝑡) (7.12b)
d𝑡𝑡
In the Laplace domain we get
1 1 𝑇𝑇d 𝑠𝑠
𝑈𝑈 𝑠𝑠 = 𝐾𝐾c 1 + 𝑅𝑅 𝑠𝑠 − 𝐾𝐾c 1 + + 𝑌𝑌(𝑠𝑠) (7.13)
𝑇𝑇i 𝑠𝑠 𝑇𝑇i 𝑠𝑠 𝑇𝑇f 𝑠𝑠+1
Process
Exercise 7.2
Control
Laboratory Which is the control law, both in the time domain and the Laplace domain, for
the series form of a PID controller with differentiation of the filtered output
measurement?
7.1.5 Setpoint weighting
A simple way of obtaining a 2DOF PID controller is to use setpoint weighting.
With the definitions
𝑒𝑒p = 𝑏𝑏𝑏𝑏 − 𝑦𝑦 , 𝑒𝑒 = 𝑟𝑟 − 𝑦𝑦 , 𝑒𝑒d = 𝑐𝑐𝑐𝑐 − 𝑦𝑦f (7.15)
where 𝑏𝑏 and 𝑐𝑐 are setpoint weights, the control law becomes
1 𝑡𝑡 d𝑒𝑒d (𝑡𝑡)
𝑢𝑢 𝑡𝑡 = 𝐾𝐾c 𝑒𝑒p 𝑡𝑡 + ∫ 𝑒𝑒 𝜏𝜏 d𝜏𝜏 + 𝑇𝑇d + 𝑢𝑢0 (7.16a)
𝑇𝑇i 0 d𝑡𝑡
d𝑦𝑦f (𝑡𝑡)
𝑇𝑇f + 𝑦𝑦f 𝑡𝑡 = 𝑦𝑦(𝑡𝑡) (7.16b)
d𝑡𝑡
Exercise 7.3
Include setpoint weighting in the series form of a PID controller.
Process
7.1.6 Non-interactive form of a PID controller
Control
In the control laws treated so far, the proportional part alone cannot be
Laboratory
Process
Control
7.2 Choice of controller type
Laboratory
The choice between controller types such as P, PI, PD, PID is considered. In
principle, the simplest controller that can do the job should be chosen.
7.2.1 On-off controller
An on-off controller is the simplest type of controller, where the control signal
has only two levels. If the variables are defined such that a positive control error
𝑒𝑒(𝑡𝑡) should be corrected by an increase of the control signal 𝑢𝑢(𝑡𝑡), the control
law is
𝑢𝑢max if 𝑒𝑒 𝑡𝑡 > 𝑒𝑒hi
𝑢𝑢 𝑡𝑡 = �𝑢𝑢0 or unchanged if 𝑒𝑒lo ≤ 𝑒𝑒 𝑡𝑡 ≤ 𝑒𝑒hi (7.22)
𝑢𝑢min if 𝑒𝑒 𝑡𝑡 < 𝑒𝑒lo
where 𝑢𝑢max , 𝑢𝑢0 , 𝑢𝑢min are the high, normal, low value of the control signal. The
interval [𝑒𝑒lo , 𝑒𝑒hi ] is a dead zone. In the simplest case, 𝑒𝑒lo = 𝑒𝑒hi = 0.
The on-off controller is inexpensive, but it causes oscillations in the pro-cess. It is
often used for temperature control in simple appliances such as ovens, irons,
refrigerators and freezers, where oscillations are tolerated.
Process
7.2.2 P controller
Control
Process
7.2.3 PI controller
Control
where the gain 𝐾𝐾c and the integral time 𝑇𝑇i are adjustable parameters; 𝑢𝑢0 is
less important due to the integral.
The main advantage of the PI controller is that there will be no remaining control
error after a setpoint change or a process disturbance. A disadvantage is that
there is a tendency for oscillations.
PI control is used when no steady-state error is desired and there is no reason to
use derivative action. Measurement noise is often a reason for not using
derivative action.
PI control is suitable for noisy processes, integrating processes and processes
resembling first-order systems. The most typical application is flow control. PI
control might also be preferable for processes with large time delays.
Process
7.2.4 PD controller
Control
d𝑒𝑒(𝑡𝑡)
𝑢𝑢 𝑡𝑡 = 𝐾𝐾c 𝑒𝑒 𝑡𝑡 + 𝑇𝑇d + 𝑢𝑢0 (7.25)
d𝑡𝑡
where the gain 𝐾𝐾c and the derivative time 𝑇𝑇d are adjustable parameters; 𝑢𝑢0 is
chosen as for a P controller.
A PD controller is preferred when integral action is not needed, but the dynamics
of the process are so slow that the predictive nature of derivative action is useful.
Many thermal processes, where energy is stored with small heat losses (e.g.,
ovens), usually have slow dynamics, almost as integrating systems. A PD
controller might then be suitable for temperature control.
Another typical application for PD control is in servo mechanisms such as
electrical motors, which usually behave as second-order integrating systems.
Process
7.2.5 PID controller
Control
As has been shown in Section 7.1, there are many variants of PID controllers.
Laboratory
The ideal form and the classical series form have 3 adjustable parameters in
addition to 𝑢𝑢0 : the proportional gain, the integral time, and the derivative
time.
If a derivative filter is included, there are 4 adjustable parameters, but the
filter time constant is usually selected as a given fraction (e.g., 10 %) of the
derivative time.
In addition, the setpoint can be weighted in the proportional part and the
derivative part.
If there is no reason to exclude integral action or derivative action, a PID
controller is the natural choice. Typically PID control is used for underdamped
processes, processes with slow dynamics and not very large time delays, and
systems of second and higher order.
Typical applications are control of temperature and chemical composition when
the process is not close to an integrating system.
Process
Control
7.3 Specifications and performance criteria
Laboratory
7.3.1 General performance criteria
The task of a controller is to control a system to behave in a desired way despite
unknown disturbances and an inaccurately known system.
The controlled system should satisfy performance criteria such as:
The controlled system must be stable; this is absolutely necessary.
The effect of disturbances on the controlled output is minimized; this is
especially important for regulatory control.
The controlled output should follow setpoint changes fast and smoothly; this
is especially important for setpoint tracking.
The control error is minimized or kept within certain limits,
The control signal variations should be moderate or at least not be excessively
large; more variations wear out control equipment faster.
The control system should be robust (insensitive) to moderate changes in
system properties, which introduce model uncertainty.
The importance of these criteria varies from case to case. Since many criteria are
conflicting, compromises have to be made in the control design.
Process
7.3.2 Fundamental limitations
Control
One reason to the fact that there are usually good solutions to the conflicting
Laboratory
The process dynamics is often the performance-limiting factor. Such factors are
Process
Control
Laboratory
time delays as well as RHP (right-half plane) poles and zeros
high-order dynamics
In practice, all processes are nonlinear. Such a process
cannot be described accurately at different operating points by a linear model
with constant parameters; thus there is model/process uncertainty.
Disturbances such as load disturbances and measurement noise limit how well a
variable can be controlled.
Efficient control of load disturbances often require derivative action, but
measurement noise is bad for the derivative.
Large load disturbances can cause the control variable to reach its (physical)
maximum or minimum value. This is especially troublesome if the controller
contains an integrator. Proportional band and integrator windup are two
concepts that deal with this limitation.
Process
7.3.3 Proportional band and integrator windup
Control
Laboratory
Proportional band
A controller’s proportional band (PB) denotes the maximum control error the
controller can handle with the available control signal. The PB is defined for a P
controller, but it can be extended to a full PID controller.
If the control signal is limited by 𝑢𝑢min ≤ 𝑢𝑢(𝑡𝑡) ≤ 𝑢𝑢max , a P controller can
according to (7.23) handle a control error that satisfies
𝑢𝑢min −𝑢𝑢0 𝑢𝑢max −𝑢𝑢0
≡ 𝑒𝑒min ≤ 𝑒𝑒(𝑡𝑡) ≤ 𝑒𝑒max ≡ (7.26)
𝐾𝐾c 𝐾𝐾c
The PB is equal to 𝑒𝑒max − 𝑒𝑒min = 𝑦𝑦hi − 𝑦𝑦lo , where 𝑦𝑦hi is the highest output
(𝑒𝑒min = 𝑟𝑟 − 𝑦𝑦hi ) and 𝑦𝑦lo is the lowest output (𝑒𝑒max = 𝑟𝑟 − 𝑦𝑦lo ) the controller
can handle. Usually, the PB is defined in percent of the total measurable output
interval 𝑦𝑦min , 𝑦𝑦max . Then, the PB is
𝑦𝑦hi −𝑦𝑦lo 𝑢𝑢max −𝑢𝑢min 100%
𝑃𝑃b = 100% = ⋅ (7.27)
𝑦𝑦max −𝑦𝑦min 𝑦𝑦max −𝑦𝑦min 𝐾𝐾c
Integrator windup
Process
Control
Laboratory
Usually controllers are tuned for stability and performance, not for signal limits.
Therefore, it is not uncommon that a control signal reaches a constraint. If the
controller contains integral action, this can be very damaging to the control
performance unless the situation is handled properly.
Consider the figure, where the PI control law (7.24) is used. A strong disturbance
causes the process output to fall well below the set-point. The controller is not
able to eliminate the control error (A)
because the control signal has reached a
constraint. During this time, the positive
control error will increase the integral in
the controller. If the disturbance later
disappears, the controller will still keep
the control signal at the constraint due
to the large value of the integral, even
If the control error goes below zero.
This will cause the output (B), which is
entirely due to the controller.
Illustration of integral windup.
There are sophisticated as well as simple methods for handling the problem. The
term anti-windup is used for such arrangements.
A simple solution is to stop integrating when a control signal reaches a constraint.
This requires that
it is known when the control signal reaches a constraint (e.g., through
measurement)
there is some built-in logic to interrupt the integration
In the case of digital control, which nowadays is customary, automatic anti-
windup can be built into the control law.
Process
7.3.4 Design specifications
Control
Error integrals
Process
Control
Laboratory
In principle, a small overshoot, rise time and settling time are desired. In
practice, the overshoot and settling time will increase with decreasing rise time,
and vice versa. Therefore, compromises have to be made.
One way of solving this problem in an optimal way is to specify some error
integral to be minimized. Examples of such error integrals are
𝑡𝑡 𝑡𝑡
𝐽𝐽IAE = ∫0 s 𝑒𝑒(𝑡𝑡) d𝑡𝑡 , 𝐽𝐽ISE = ∫0 s 𝑒𝑒(𝑡𝑡)2 d𝑡𝑡
𝑡𝑡s 𝑡𝑡s (7.30)
𝐽𝐽ITAE = ∫0 𝑡𝑡 𝑒𝑒(𝑡𝑡) d𝑡𝑡 , 𝐽𝐽ITSE = ∫0 𝑡𝑡𝑒𝑒(𝑡𝑡)2 d𝑡𝑡
where the acronyms are
– IAE = “integrated absolute error”
– ISE = “integrated square error”
– ITAE = “integrated time-weighted absolute error”
– ITSE = “integrated time-weighted square error”
The weighting with time forces the control error towards zero as time increases.
In principle, the integration time should be infinite, but because the minimization
has to be done numerically, a finite 𝑡𝑡s has to be used.
In the figure, IAE and ISE are normalized with 𝜔𝜔n , ITAE and ITSE with 𝜔𝜔n2 . As can
be seen, every normalized error integral has a minimum for a given relative
damping 𝜁𝜁 . This damping as well
as the corresponding relative
overshoot 𝑀𝑀 are shown below.
Table 7.1 Optimal relative
damping for 2nd order system.
Process
Control
7.4 Tuning based on frequency response
Laboratory
7.4.1 Experimental tuning
An ideal PID controller of interactive
form can be tuned experimentally G
by making closed-loop control experi-
ments with the real process. The
standard feedback structure is used.
1. A P controller (𝐺𝐺c = 𝐾𝐾c ) is used for the first experiment. A low value is chosen
for the gain 𝐾𝐾c . Note that 𝐾𝐾c must have the same sign as 𝐾𝐾p .
2. A change in the setpoint 𝑅𝑅 is introduced. (Some other disturbance could also
be used.) The controller gain 𝐾𝐾c is increased until the output 𝑌𝑌 starts to
oscillate with a constant amplitude (see next slide).
3. The value of the controller gain yielding constant oscillations is denoted
𝐾𝐾c,max . The period of the oscillations is denoted 𝑃𝑃c .
4. The controller gain is changed to 𝐾𝐾c = 0.5𝐾𝐾c,max . If the intention was to tune
a P controller, this is the final tuning.
The next figure shows how changes of the controller gain 𝐾𝐾c and the integral time
𝑇𝑇i typically affect the control performance. The optimal performance is in this
case obtained by 𝐾𝐾c = 3 and 𝑇𝑇i = 11 .
𝑇𝑇i = 5 𝑇𝑇i = 11 𝑇𝑇i = 20
𝐾𝐾c = 5
𝐾𝐾c = 3
𝐾𝐾c = 1
Process
7.4.2 Ziegler-Nichols’s recommendations
Control
In 1942, Ziegler and Nichols suggested tunings for P, PI and PID controllers based
Laboratory
Process
7.4.3 Åström’s and Hägglund’s correlations
Control
In 2006, Åström and Hägglund showed that, in general, 𝐾𝐾c,max and 𝑃𝑃c alone do
Laboratory
Process
Control
7.5 Tuning based on step response
Laboratory
A drawback with generating the frequency response is that it is quite
cumbersome and time-consuming to generate oscillations with constant
amplitude by adjusting a controller parameter.
An alternative is to use a step response for the process.
The figure illustrates how the
needed parameters are obtained
from a unit-step response, i.e., a
step with size 𝑢𝑢step = 1 expressed
in the units used for the control
variable. yi
Instead of taking the 𝑎𝑎 parameter from the point, where the tangent through
Process
Control the inflexion point (i.e., the point where the slope is highest) of the step response
Laboratory
crosses the vertical axis, it can be calculated when the coordinates (𝑡𝑡i , 𝑦𝑦i ) of the
inflexion point are known. The calculation is valid for any size of 𝑢𝑢step . The
formula for 𝑎𝑎 is
𝐿𝐿𝑦𝑦i
𝑎𝑎 = (7.32)
𝑢𝑢step (𝑡𝑡i −𝐿𝐿)
Process
7.5.1 Ziegler-Nichols’s recommendations
Control
In 1942, Ziegler and Nichols suggested tunings for P, PI and PID controllers
Laboratory
based also on the information that can be obtained from a step test. Their
recommendations for an ideal controller are given in Table 7.4.
The method requires 𝐿𝐿 > 0 and preferably 0.1 ≤ 𝜃𝜃 ≤ 1.
Controller aK c Ti / L Td / L
P 1.0 – –
PI 0.9 3 –
PID 1.2 2 0.5
Process
7.5.2 The CHR method
Control
Table 7.5. Controller tuning for regulatory control by the CHR method.
Process
Control
Laboratory
No overshoot 20 % overshoot
Controller
aK c Ti / L Td / L aK c Ti / L Td / L
P 0.3 – – 0.7 – –
PI 0.6 4.0 – 0.7 2.3 –
PID 0.95 2.4 0.42 1.2 2.0 0.42
Table 7.6. Controller tuning for setpoint tracking by the CHR method.
No overshoot 20 % overshoot
Controller
aK c TTi / Teq Td / L aK c TTii //TTeq Td / L
P 0,3 – – 0,7 – –
PI 0,35 1,2 – 0,6 1,0 –
PID 0,6 1,0 0,5 0,95 1,4 0,47
Process
7.5.3 Åström’s and Hägglund’s correlations
Control
θ 13
PI 0.35 + 0.15θ − 0.35 + –
(1 + θ ) 2 1 + 12 θ + 7θ 2
8 + 4θ 0.5
PID 0.45 + 0.2 θ
1 + 10 θ 1 + 0.3θ
Process
Control
7.6 Model-based controller tuning
Laboratory
The controller tuning methods in Sections 7.4 and 7.5 employ parameters that
can be determined from an experiment with an existing process.
If a process model is known, the same parameters can be determined
through a simulation experiment
possibly by direct calculation from the process model
For example, a first-order system with a time delay has the transfer function
𝐾𝐾
𝐺𝐺 𝑠𝑠 = e−𝐿𝐿𝐿𝐿 (7.34)
𝑇𝑇𝑇𝑇+1
The same tuning methods as in Sections 7.4 and 7.5 can then be used.
However, the methods in Sections 7.4 and 7.5 are “general purpose” methods
that are not optimized for any specific model type.
For a given model, better controller tunings probably exist.
KEH Process Dynamics and Control 7–39
7. PID Controllers 7.6 Model-based controller tuning
Process
7.6.1 First-order system with a time delay
Control
IAE 0.902 θ −0.985 0.984 θ −0.986 1.645 θ 0.707 1.435 θ −0.921 1.139 θ 0.749 0.482 θ 1.137
ITAE 0.490 θ −1.084 0.859 θ −0.977 1.484 θ 0.680 1.357 θ −0.947 1.188 θ 0.738 0.381 θ 0.995
Table 7.9. IAE and ITAE minimizing controller tunings for setpoint tracking.
IAE 0.758 θ −0.861 (1.020 − 0.323θ ) −1 1.086 θ −0.869 (0.740 − 0.130 θ ) −1 0.348 θ 0.914
ITAE 0.586 θ −0.916 (1.030 − 0.165θ ) −1 0.965 θ −0.855 (0.796 − 0.147 θ ) −1 0.308 θ 0.929
Table 7.10. Cvejn’s tunings for regulatory control and setpoint tracking.
PI controller PID controller
Control
KK c Ti / T KK c Ti / T Td / T
The PI controller tunings tend to give better robustness than the PID controller
tunings, which tend to give better performance.
Process
7.6.2 Second-order no-zero system with a time delay
Control
We shall consider second-order systems with a time delay but no zeros. Such a
Laboratory
In Cvejn’s method for tracking control, the controller 𝐺𝐺c (𝑠𝑠) is tuned to give the
loop transfer 𝐺𝐺ℓ (𝑠𝑠) = 𝐺𝐺(𝑠𝑠)𝐺𝐺c (𝑠𝑠) such that
1 −𝐿𝐿𝐿𝐿
𝐺𝐺ℓ 𝑠𝑠 = e (7.37)
2𝐿𝐿𝐿𝐿
or
1 3
𝐺𝐺ℓ 𝑠𝑠 = 1+ e−𝐿𝐿𝐿𝐿 (7.38)
4 𝐿𝐿𝐿𝐿
Exercise 7.3
Use Cvejn’s method for tracking control to tune a PID controller for the system
(7.36).
Cvejn’s method can be used also in this case, but Åström and Hägglund (2006)
suggest the following tuning when the system is overdamped:
𝐾𝐾𝐾𝐾c = 0.19 + 0.37𝜃𝜃1−1 + 0.18𝜃𝜃2−1 + 0.02𝜃𝜃1−1 𝜃𝜃2−1
𝐾𝐾𝐾𝐾c 𝐿𝐿/𝑇𝑇i = 0.48 + 0.03𝜃𝜃1−1 − 0.0007𝜃𝜃2−1 + 0.0012𝜃𝜃1−1 𝜃𝜃2−1 (7.40)
𝜃𝜃1 +𝜃𝜃2
𝐾𝐾𝐾𝐾c 𝑇𝑇d /𝐿𝐿 = 0.29 + 0.16𝜃𝜃1−1 − 0.2𝜃𝜃2−1 + 0.28𝜃𝜃1−1 𝜃𝜃2−1
𝜃𝜃1 +𝜃𝜃2 +𝜃𝜃1 𝜃𝜃2
where
𝜃𝜃1 = 𝐿𝐿/𝑇𝑇1 , 𝜃𝜃2 = 𝐿𝐿/𝑇𝑇2 (7.41)
For this kind of system, Åström and Hägglund (2006) suggest the tuning:
𝐾𝐾𝐾𝐾c 𝐿𝐿 = 0.37 + 0.02𝜃𝜃2−1
𝐾𝐾𝐾𝐾c 𝐿𝐿2 /𝑇𝑇i = 0.03 + 0.0012𝜃𝜃2−1 (7.43)
𝐾𝐾𝐾𝐾c 𝑇𝑇d = 0.16 + 0.28𝜃𝜃2−1
Process
Control
7.7 Controller design by direct synthesis
Laboratory
In the previous sections, equations for controller tuning have been given for first-
and second-order no-zero systems.
The equations are usually the result of optimization of some criterion that is
considered to imply “good control”.
However, what is “good control” varies from case to case depending on the
compromise between stability and performance.
A drawback of the tuning equations is that the user cannot influence the
tuning according to his/her opinion of “good control”.
In this section, a method is introduced whereby
the user can influence the controller tuning in a systematic way according to
his/her opinion of “good control”
more model types than in previous sections can be handled, e.g., systems with
a zero
Process
7.7.1 Closed-loop transfer functions
Control
Process
7.7.2 Low-order minimum-phase systems
Control
Laboratory
First-order system
A strictly proper first-order system without a time delay has the transfer function
𝐾𝐾
𝐺𝐺 = (7.52)
𝑇𝑇𝑇𝑇+1
Assume that we want the controlled system to behave as a first-order system
with the time constant 𝑇𝑇r . Then,
1 𝐺𝐺r 1
𝐺𝐺r = , which gives = (7.53)
𝑇𝑇r 𝑠𝑠+1 1−𝐺𝐺r 𝑇𝑇r 𝑠𝑠
Substitution of (7.52) and (7.53) into (7.51) gives
𝑇𝑇𝑇𝑇+1 1 𝑇𝑇 1
𝐺𝐺c = = 1+ (7.54)
𝐾𝐾 𝑇𝑇r 𝑠𝑠 𝐾𝐾𝑇𝑇r 𝑇𝑇𝑇𝑇
Here, 𝑇𝑇r is a design parameter, by which the performance of the control system
can be affected.
Even if the uncontrolled system is of second order, we can specify the controlled
system to be of first order. Substitution of (7.53) and (7.56) into (7.51) then gives
𝑠𝑠 2 +2𝜁𝜁𝜔𝜔n 𝑠𝑠+𝜔𝜔n
2 1 2𝜁𝜁 𝜔𝜔n 𝑠𝑠
𝐺𝐺c = 2 = 1+ + (7.57)
𝐾𝐾𝜔𝜔n 𝑇𝑇r 𝑠𝑠 𝐾𝐾𝜔𝜔n 𝑇𝑇r 2𝜁𝜁𝑠𝑠 2𝜁𝜁𝜔𝜔n
Also here, 𝑇𝑇r is a design parameter which only affects the controller gain.
Process
7.7.3 High-order minimum-phase systems
Control
A high-order minimum-phase system with real poles and zeros, but with no time
Laboratory
Process
7.7.4 Second-order system with RHP zero
Control
A second-order system with real poles and a right half-plane (RHP) zero has the
Laboratory
transfer function
𝐾𝐾(−𝑇𝑇3 𝑠𝑠+1)
𝐺𝐺 𝑠𝑠 = , 𝑇𝑇𝑖𝑖 ≥ 0 (7.63)
(𝑇𝑇1 𝑠𝑠+1)(𝑇𝑇2 𝑠𝑠+1)
Now division by 𝐺𝐺 in (7.51) will result in an unstable controller with a RHP pole if
𝐺𝐺r is chosen as in the previous sections.
One possible solution is to approximate the unstable controller by a stable
controller. This tends to result in too aggressive control because the
controller is then designed as if there were no RHP zero in 𝐺𝐺 .
Another solution is to include the same RHP zero in 𝐺𝐺r as in 𝐺𝐺 ; it will then be
cancelled out in (7.51) and the controller will automatically be stable. This
means that the choice of 𝐺𝐺r is restricted, but otherwise the control
performance tends to be as expected.
Analogously with the derivation of (7.61), this gives the PID controller parameters
𝑇𝑇f 𝜔𝜔r2 𝑇𝑇1 𝑇𝑇2
𝐾𝐾c = (𝑇𝑇1 + 𝑇𝑇2 − 𝑇𝑇f ), 𝑇𝑇i = 𝑇𝑇1 + 𝑇𝑇2 − 𝑇𝑇f , 𝑇𝑇d = − 𝑇𝑇f (7.70)
𝐾𝐾 𝑇𝑇1 +𝑇𝑇2 −𝑇𝑇f
where 𝑇𝑇f , given by (7.68), is the derivative filter time constant in a PID controller
(7.60).
Let 𝐺𝐺r have real poles at −1/𝑇𝑇r and −1/𝑇𝑇3 . This corresponds to
𝜁𝜁r = 0.5(𝑇𝑇r + 𝑇𝑇3 )𝜔𝜔r and 𝜔𝜔r = 1/ 𝑇𝑇r 𝑇𝑇3 , which for (7.68) gives
𝑇𝑇r 𝑇𝑇3
𝑇𝑇f = (7.72)
𝑇𝑇r +2𝑇𝑇3
Process
7.7.5 First-order system with a time delay
Control
To illustrate how systems with a time delay can be handled by direct synthesis, a
Laboratory
first-order system with a time delay will be studied. Such a system has the
transfer function
𝐾𝐾
𝐺𝐺 𝑠𝑠 = e−𝐿𝐿𝐿𝐿 (7.73)
𝑇𝑇𝑠𝑠+1
Substitution of (7.75) and (7.76) into (7.51) gives a PID controller with the
parameters
𝑇𝑇+0.5𝐿𝐿−𝑇𝑇f 0.5𝐿𝐿𝐿𝐿 0.5𝐿𝐿𝑇𝑇r
𝐾𝐾c = , 𝑇𝑇i = 𝑇𝑇 + 0.5𝐿𝐿 − 𝑇𝑇f , 𝑇𝑇d = , 𝑇𝑇f = (7.77)
𝐾𝐾(𝑇𝑇r +𝐿𝐿) 𝑇𝑇+0.5𝐿𝐿−𝑇𝑇f 𝑇𝑇r +𝐿𝐿
Here, 𝑇𝑇f is the time constant of a derivative filter in the PID controller (7.60).
Process
Control
7.8 Internal model control
Laboratory
“Internal model control” (IMC) is closely related to “direct synthesis” (DS). As in
DS, a model of the system to be controlled is explicitly built into the controller,
but in a different way.
An advantage with IMC is that it is easier to implement more complex control
laws than regular PID controllers. For example, the controller transfer function
(7.79) can easily be implemented exactly with IMC.
Even if the controller design is based on IMC, it is often desirable to
implement the controller as a regular PID controller. In such cases, the IMC
approach offers better possibilities to deal with robustness issues than DS.
Process
7.8.1 The IMC structure
Control
Process
7.8.2 Handling of time delays without approximation
Control
𝐺𝐺� = 𝐾𝐾e−𝐿𝐿𝐿𝐿 /(𝑇𝑇𝑇𝑇 + 1). Choose the IMC filter as 𝐺𝐺f = e−𝐿𝐿𝐿𝐿 /(𝑇𝑇r 𝑠𝑠 + 1) . Substitution
into (7.82) now gives
1 𝑇𝑇𝑇𝑇+1 1 𝑇𝑇−𝑇𝑇r
𝐺𝐺IMC = = 1+ 𝑠𝑠 (7.84)
𝐾𝐾 𝑇𝑇r 𝑠𝑠+1 𝐾𝐾 𝑇𝑇r 𝑠𝑠+1
which is a PD controller with a derivative filter having the parameters 𝐾𝐾𝑐𝑐 = 1/𝐾𝐾 ,
𝑇𝑇d = 𝑇𝑇 − 𝑇𝑇r , 𝑇𝑇f = 𝑇𝑇r . Substitution of (7.84) and the model 𝐺𝐺� into (7.81) gives
𝑇𝑇𝑇𝑇+1
𝐺𝐺c = (7.85)
𝐾𝐾(𝑇𝑇r 𝑠𝑠+1−e−𝐿𝐿𝐿𝐿 )
which is identical with (7.79). The difference is that (7.85) can be implemented
exactly by the IMC structure without time-delay approximation.
Note that there is no integration in 𝐺𝐺IMC , but the feedback of 𝐺𝐺� in the IMC
structure introduces integration if 𝐺𝐺IMC is calculated using the same 𝐺𝐺� in
(7.82); integration is achieved even if 𝐺𝐺� ≠ 𝐺𝐺 .
Exercise. Calculate the closed-loop transfer function 𝐺𝐺r when 𝐺𝐺� ≠ 𝐺𝐺 . Show
that there will be no steady-state error, i.e., that 𝐺𝐺r 0 = 1 .
Process
7.8.3 The predictive character of the IMC structure
Control
The previous block diagram of the IMC structure is drawn to emphasize how
Laboratory
Process
7.8.4 Controller design
Control
Process
7.8.5 Implementation with a regular PID controller
Control
An advantage of the IMC structure is that time delays can be handled exactly, but
Laboratory
K e− Ls
Ti / λ T1 + 12 L 1 LT1 / Ti Tr + 12 L
T1s + 1 2
K (T3 s + 1) e− Ls
Ti / λ T1 + T2 − T3 (T1T2 / Ti ) − T3 Tr + L
(T1s + 1)(T2 s + 1)
K (−T3 s + 1) e− Ls
Ti / λ T1 + T2 + (T3 L / λ ) (T1T2 / Ti ) − (T3 L / λ ) Tr + T3 + L
(T1s + 1)(T2 s + 1)
K e− Ls L(1 − 12 L / Ti ) Tr + 12 L
Ti / λ 2 2λ 1
2
s
K e− Ls
Ti / λ 2 2λ + T2 − L T2 (1 − T2 / Ti ) Tr + L
s (T2 s + 1)
The desired time constant of the closed-loop system is 𝑇𝑇r . 𝜆𝜆 , which is used in
the calculations, is closely related to 𝑇𝑇r . Note that the calculated integral time
𝑇𝑇i is used in several expressions.
Process
Control
7.9 Model simplification
Laboratory
Many controller tuning methods have been presented in the previous sections.
Section 7.4: Controller tuning based on frequency-response parameters
𝐾𝐾c,max , 𝑃𝑃c (or 𝜔𝜔c ) and 𝜅𝜅. These methods are “general-purpose methods”
not optimized for any specific model type.
Section 7.5: Controller tuning based on step-response parameters 𝑎𝑎 (or 𝑡𝑡i , 𝑦𝑦i ),
𝐿𝐿 and 𝜃𝜃. These methods are also general-purpose methods not optimized for
any specific model type.
Section 7.6: Model-based tuning optimized for given model structures and
control criteria with no user interaction.
Section 7.7: Direct synthesis for low-order models according to desired closed-
loop response.
Section 7.8: Internal model control mainly for low-order models according to
desired closed-loop response.
In this section, methods to reduce high-order models to first- or second-order
models are presented. Any controller tuning method can be used.
Process
7.9.1 Skogestad’s method
Control
model with real poles and zeros to a first- or second-order model with a time
delay but with no zeros.
The transfer function to be simplified is factorized into a minimum-phase part
𝐺𝐺 ⊖ and a non-minimum-phase part 𝐺𝐺 ⊕ , i.e.,
𝐺𝐺 𝑠𝑠 = 𝐺𝐺 ⊕ (𝑠𝑠)𝐺𝐺 ⊖ (s) (7.92)
Any left-half plane (LHP) zeros of 𝐺𝐺 ⊖ (s) and RHP zeros of 𝐺𝐺 ⊕ (𝑠𝑠) are eliminated
by suitable approximations.
Elimination of LHP zeros
If the poles and zeros are real, the minimum-phase part has the form
𝐾𝐾 𝑇𝑇𝑛𝑛+1 𝑠𝑠+1 𝑇𝑇𝑛𝑛+2 𝑠𝑠+1 …(𝑇𝑇𝑛𝑛+𝑚𝑚 𝑠𝑠+1)
𝐺𝐺 ⊖ 𝑠𝑠 = (7.93)
𝑇𝑇1 𝑠𝑠+1 𝑇𝑇2 𝑠𝑠+1 …(𝑇𝑇𝑛𝑛 𝑠𝑠+1)
where 𝑇𝑇1 ≥ 𝑇𝑇2 ≥ ⋯ ≥ 𝑇𝑇𝑛𝑛 > 0, 𝑇𝑇𝑛𝑛+1 ≥ 𝑇𝑇𝑛𝑛+2 ≥ ⋯ ≥ 𝑇𝑇𝑛𝑛+𝑚𝑚 > 0 , 𝑛𝑛 > 𝑚𝑚. The
simplification procedure now goes as follows.
The numerator time constants 𝑇𝑇𝑛𝑛+1 , 𝑇𝑇𝑛𝑛+2 , …, 𝑇𝑇𝑛𝑛+𝑚𝑚 are considered in that
order. Assume that 𝑇𝑇𝑛𝑛+𝑗𝑗 is the one currently being considered.
Next, the smallest remaining denominator time constant 𝑇𝑇𝑖𝑖 such that
Process
Control 𝑇𝑇𝑖𝑖 ≥ 𝑇𝑇𝑛𝑛+𝑗𝑗 is selected. If there is no such time constant, or if 𝑇𝑇𝑖𝑖 ≫ 𝑇𝑇𝑛𝑛+𝑗𝑗 ,
Laboratory
the smaller 𝑇𝑇𝑖𝑖 closest to 𝑇𝑇𝑛𝑛+𝑗𝑗 is chosen. It is considered that 𝑇𝑇𝑖𝑖 ≫ 𝑇𝑇𝑛𝑛+𝑗𝑗
2
if 𝑇𝑇𝑖𝑖 > 𝑇𝑇𝑛𝑛+𝑗𝑗 /𝑇𝑇𝑖𝑖+1 and 𝑇𝑇𝑛𝑛+𝑗𝑗 /𝑇𝑇𝑖𝑖+1 < 1.6 .
The ratio (𝑇𝑇𝑛𝑛+𝑗𝑗 𝑠𝑠 + 1)/(𝑇𝑇𝑖𝑖 𝑠𝑠 + 1) is now approximated as
𝑇𝑇𝑛𝑛+𝑗𝑗 /𝑇𝑇𝑖𝑖 if 𝑇𝑇𝑖𝑖 ≥ 𝑇𝑇𝑛𝑛+𝑗𝑗 ≥ 5𝑇𝑇r a
5𝑇𝑇r /𝑇𝑇𝑖𝑖
if 𝑇𝑇𝑖𝑖 ≥ 5𝑇𝑇r ≥ 𝑇𝑇𝑛𝑛+𝑗𝑗 b
5𝑇𝑇r −𝑇𝑇𝑛𝑛+𝑗𝑗 𝑠𝑠+1
𝑇𝑇𝑛𝑛+𝑗𝑗 𝑠𝑠+1 1
≈ if 5𝑇𝑇r ≥ 𝑇𝑇𝑖𝑖 ≥ 𝑇𝑇𝑛𝑛+𝑗𝑗 c (7.94)
𝑇𝑇𝑖𝑖 −𝑇𝑇𝑛𝑛+𝑗𝑗 𝑠𝑠+1
𝑇𝑇𝑖𝑖 𝑠𝑠+1
𝑇𝑇𝑛𝑛+𝑗𝑗 /𝑇𝑇𝑖𝑖 if 𝑇𝑇𝑛𝑛+𝑗𝑗 ≥ 𝑇𝑇𝑖𝑖 ≥ 𝑇𝑇r (d)
𝑇𝑇𝑛𝑛+𝑗𝑗 /𝑇𝑇r if 𝑇𝑇𝑛𝑛+𝑗𝑗 ≥ 𝑇𝑇r ≥ 𝑇𝑇𝑖𝑖 (e)
1 if 𝑇𝑇r ≥ 𝑇𝑇𝑛𝑛+𝑗𝑗 ≥ 𝑇𝑇𝑖𝑖 (f)
Here, 𝑇𝑇r is the desired closed-loop time constant. If this is not known, the
suggested value is 𝑇𝑇r = 𝐿𝐿� , which is the time delay in the simplified model.
Since this is not initially known, one may have to iterate (i.e., first guessing 𝐿𝐿� ,
then possibly correcting with the new 𝐿𝐿� ).
Note that the gain as well as the values and number of denominator time
constants may have changed from the original 𝐺𝐺 ⊖ .
Process
7.9.2 Isaksson’s and Graebe’s method
Control
model, where the fast and slow dynamics are combined to yield a model with a
desired number of poles and zeros. If the original model contains a time delay, it
is either left intact or substituted by a Padé approximation.
To describe the method, both factorized and polynomial forms of the original
transfer function are employed. If the numerator order is 𝑚𝑚 and the
denominator order is 𝑛𝑛 , the transfer function is
𝑇𝑇𝑛𝑛+1 𝑠𝑠+1 𝑇𝑇𝑛𝑛+2 𝑠𝑠+1 …(𝑇𝑇𝑛𝑛+𝑚𝑚 𝑠𝑠+1)
𝐺𝐺 𝑠𝑠 = 𝐾𝐾 (7.99a)
𝑇𝑇1 𝑠𝑠+1 𝑇𝑇2 𝑠𝑠+1 …(𝑇𝑇𝑛𝑛 𝑠𝑠+1)
𝑏𝑏0 𝑠𝑠 𝑚𝑚 +⋯+𝑏𝑏𝑚𝑚−2 𝑠𝑠 2 +𝑏𝑏𝑚𝑚−1 𝑠𝑠+1
= 𝐾𝐾 (7.99b)
𝑎𝑎0 𝑠𝑠 𝑛𝑛 +⋯+𝑎𝑎𝑛𝑛−2 𝑠𝑠 2 +𝑎𝑎𝑛𝑛−1 𝑠𝑠+1
where 𝑇𝑇1 ≥ 𝑇𝑇2 ≥ ⋯ ≥ 𝑇𝑇𝑛𝑛 > 0 (i.e., a stable system) and |𝑇𝑇𝑛𝑛+1 | ≥ |𝑇𝑇𝑛𝑛+2 | ≥
⋯ ≥ |𝑇𝑇𝑛𝑛+𝑚𝑚 | . The numerator time constants can be positive or negative.
If the model is to be used for controller tuning, a strictly proper first- or second-
order model, possibly with a time delay, is usually desired. Then
𝐾𝐾
𝐺𝐺� 𝑠𝑠 = 1 (1st order) (7.101)
𝑇𝑇1 +𝑎𝑎𝑛𝑛−1 𝑠𝑠+1
2
1
𝐾𝐾 𝑇𝑇𝑛𝑛+1 +𝑏𝑏𝑚𝑚−1 𝑠𝑠+1
𝐺𝐺� 𝑠𝑠 = 1 2
1 (2nd order) (7.102)
𝑇𝑇1 𝑇𝑇2 +𝑎𝑎𝑛𝑛−2 𝑠𝑠 2 + 𝑇𝑇1 +𝑇𝑇2 +𝑎𝑎𝑛𝑛−1 𝑠𝑠+1
2 2
where
𝑏𝑏𝑚𝑚−1 = ∑𝑚𝑚 𝑛𝑛
𝑗𝑗=1 𝑇𝑇𝑛𝑛+𝑗𝑗 , 𝑎𝑎𝑛𝑛−1 = ∑𝑖𝑖=1 𝑇𝑇𝑖𝑖 , 𝑎𝑎𝑛𝑛−2 = 2
1
∑𝑛𝑛
2 𝑛𝑛 2
𝑖𝑖=1 𝑇𝑇𝑖𝑖 −∑𝑖𝑖=1 𝑇𝑇𝑖𝑖 (7.103)
This model has complex-conjugated poles, but according to (7.91), 𝑇𝑇1 + 𝑇𝑇2 = 81
and 𝑇𝑇1 𝑇𝑇2 = 1843.5 can be used in the controller calculations. Table 7.12 for
IMC-based tuning of second-order model then gives
– 𝜆𝜆 = 𝑇𝑇r + 𝐿𝐿 = 10 + 2 = 12
– 𝑇𝑇i = 𝑇𝑇�1 + 𝑇𝑇�2 − 𝑇𝑇�3 = 81 − 14 = 67
– 𝐾𝐾c = 𝑇𝑇i /(𝐾𝐾𝜆𝜆) = 67/(1 ⋅ 12) = 5.6 (much bigger than in Ex. 7.2!)
– 𝑇𝑇d = 𝑇𝑇�1 𝑇𝑇�2 /𝑇𝑇i − 𝑇𝑇�3 = 1843.5/67 −14 = 13.5