Chapter 12

Controller Tuning: A Motivational Example

Fig. 12.1. Unit-step disturbance responses for the candidate controllers

(FOPTD Model: K = 1, 4, 20).

PID Controller Design, Tuning, and

Troubleshooting
Chapter 12

Performance Criteria For Closed-Loop Systems

The function of a feedback control system is to ensure that
the closed loop system has desirable dynamic and steadystate response characteristics.
Ideally, we would like the closed-loop system to satisfy the
following performance criteria:
1. The closed-loop system must be stable.
2. The effects of disturbances are minimized, providing
good disturbance rejection.
3. Rapid, smooth responses to set-point changes are
obtained, that is, good set-point tracking.

4. Steady-state error (offset) is eliminated.

Chapter 12

5. Excessive control action is avoided.

6. The control system is robust, that is, insensitive to
changes in process conditions and to inaccuracies in the
process model.
PID controller settings can be determined by a number
of alternative techniques:
1. Direct Synthesis (DS) method
2. Internal Model Control (IMC) method
3. Controller tuning relations
4. Frequency response techniques
5. Computer simulation
6. On-line tuning after the control system is installed.

Direct Synthesis Method

Chapter 12

In the Direct Synthesis (DS) method, the controller design is

based on a process model and a desired closed-loop transfer
function.
The latter is usually specified for set-point changes, but
responses to disturbances can also be utilized (Chen and
Seborg, 2002).
Although these feedback controllers do not always have a PID
structure, the DS method does produce PI or PID controllers
for common process models.
As a starting point for the analysis, consider the block diagram
of a feedback control system in Figure 12.2. The closed-loop
transfer function for set-point changes was derived in Section
11.2:
K mGcGvG p
Y

(12-1)
Ysp 1 GcGv G p Gm
4

Chapter 12
Fig. 12.2. Block diagram for a standard feedback control system.

Chapter 12

For simplicity, let G @GvG p Gm and assume that Gm = Km. Then

Eq. 12-1 reduces to
GcG
Y

(12-2)
Ysp 1 Gc G
Rearranging and solving for Gc gives an expression for the
feedback controller:
1 Y / Ysp
Gc
(12-3a)

G 1 Y / Ysp
Equation 12-3a cannot be used for controller design because
the closed-loop transfer function Y/Ysp is not known a priori.
Also, it is useful to distinguish between the actual process G
%
and the model, G , that provides an approximation of the
process behavior.
A practical design equation can be derived by replacing the
%
unknown G by G, and Y/Ysp by a desired closed-loop transfer
6

Chapter 12

1 Y / Ysp d
Gc
(12-3b)
%
G 1 Y / Ysp
d

The specification of (Y/Ysp)d is the key design decision and will

be considered later in this section.

Note that the controller transfer function in (12-3b) contains

the inverse of the process model owing to the 1/ G% term.
This feature is a distinguishing characteristic of model-based
control.

Desired Closed-Loop Transfer Function

For processes without time delays, the first-order model in
Eq. 12-4 is a reasonable choice,
Y
1

(12-4)

Ysp d c s 1

 The model has a settling time of ~ 4c , as shown in

Section 5. 2.

Chapter 12

Because the steady-state gain is one, no offset occurs for setpoint changes.
By substituting (12-4) into (12-3b) and solving for Gc, the
controller design equation becomes:
Gc

1 1
G% c s

(12-5)

The 1/ c s term provides integral control action and thus

eliminates offset.
Design parameter c provides a convenient controller tuning
parameter that can be used to make the controller more
aggressive (small c ) or less aggressive (large c ).
8

Chapter 12

If the process transfer function contains a known time delay ,

a reasonable choice for the desired closed-loop transfer
function is:
Y

Y
sp

e s

c s 1

(12-6)

The time-delay term in (12-6) is essential because it is

physically impossible for the controlled variable to respond to
a set-point change at t = 0, before t = .
If the time delay is unknown, must be replaced by an
estimate.
Combining Eqs. 12-6 and 12-3b gives:
1
e s
Gc
G%c s 1 e s

(12-7)
9

 Although this controller is not in a standard PID form, it is

physically realizable.

Chapter 12

Next, we show that the design equation in Eq. 12-7 can be used
to derive PID controllers for simple process models.
The following derivation is based on approximating the timedelay term in the denominator of (12-7) with a truncated Taylor
series expansion:
es 1 s

(12-8)

Substituting (12-8) into the denominator of Eq. 12-7 and

rearranging gives
1 e s
Gc
s
G% c

(12-9)

Note that this controller also contains integral control action.

10

First-Order-plus-Time-Delay (FOPTD) Model

Consider the standard FOPTD model,

Chapter 12

s
Ke
G% s
s 1

(12-10)

Substituting Eq. 12-10 into Eq. 12-9 and rearranging gives a PI

controller, Gc K c 1 1/ I s , with the following controller
settings:
1
Kc
,
I
(12-11)
K c

Second-Order-plus-Time-Delay (SOPTD) Model

Consider a SOPTD model,
s
Ke
G% s
1s 1 2 s 1

(12-12)
11

Substitution into Eq. 12-9 and rearrangement gives a PID

controller in parallel form,

Chapter 12

1
Gc K c 1

I s

D s

(12-13)

where:
1 1 2
Kc
,
K c

I 1 , 2

1 2

1 2

(12-14)

Example 12.1
Use the DS design method to calculate PID controller settings for
the process:
2e s
G
10s 1 5s 1
12

Chapter 12

Consider three values of the desired closed-loop time constant:

. Evaluate the controllers for unit step
c 1, 3, and 10
changes in both the set point and the disturbance, assuming that
Gd = G. Repeat the evaluation for two cases:
a. The process model is perfect ( G%= G).
b. The model gain is K%= 0.9, instead of the actual value, K = 2.
Thus,

s
0.9
e
G%
10s 1 5s 1

The controller settings for this example are:

Kc
Kc
I
D

K% 2
K% 0.9

c 1
3.75

c 3
1.88

8.33

4.17

15
3.33

15
3.33

c 10
0.682
1.51
15
3.33
13

Chapter 12

The values of Kc decrease as c increases, but the values of I

and D do not change, as indicated by Eq. 12-14.

Figure 12.3 Simulation results for Example 12.1 (a): correct

model gain.
14

Chapter 12
Fig. 12.4 Simulation results for Example 12.1 (b): incorrect
model gain.

15

Internal Model Control (IMC)

Chapter 12

A more comprehensive model-based design method, Internal

Model Control (IMC), was developed by Morari and
coworkers (Garcia and Morari, 1982; Rivera et al., 1986).
The IMC method, like the DS method, is based on an assumed
process model and leads to analytical expressions for the
controller settings.
These two design methods are closely related and produce
identical controllers if the design parameters are specified in a
consistent manner.
The IMC method is based on the simplified block diagram
shown in Fig. 12.6b. A process model G%and the controller
output P are used to calculate the model response, Y%.
16

Chapter 12

Figure 12.6.
Feedback control
strategies

The model response is subtracted from the actual response Y,

and the difference, Y Y%is used as the input signal to the IMC
controller, Gc*.
In general, Y Y%due to modeling errors G% G and unknown
disturbances D 0 that are not accounted for in the model.

The block diagrams for conventional feedback control and

IMC are compared in Fig. 12.6.
17

 It can be shown that the two block diagrams are identical if

controllers Gc and Gc* satisfy the relation

Chapter 12

Gc

Gc*

(12-16)

1 Gc*G%

Thus, any IMC controller Gc* is equivalent to a standard

feedback controller Gc, and vice versa.
The following closed-loop relation for IMC can be derived from
Fig. 12.6b using the block diagram algebra of Chapter 11:
1 Gc*G%
Y
Ysp
D
*
*
1 Gc G G%
1 Gc G G%
Gc*G

(12-17)

18

For the special case of a perfect model, G% G , (12-17) reduces to

Y Gc*GYsp 1 Gc*G D

(12-18)

Chapter 12

The IMC controller is designed in two steps:

Step 1. The process model is factored as
%
G% G%
G

(12-19)

where G%
contains any time delays and right-half plane
zeros.
In addition, G% is required to have a steady-state gain equal

to one in order to ensure that the two factors in Eq. 12-19

are unique.

19

Step 2. The controller is specified as

Chapter 12

Gc*

f
%
G

(12-20)

where f is a low-pass filter with a steady-state gain of one. It

typically has the form:
f

c s 1

(12-21)

In analogy with the DS method, c is the desired closed-loop time

constant. Parameter r is a positive integer. The usual choice is
r = 1.

20

For the ideal situation where the process model is perfect G% G ,

substituting Eq. 12-20 into (12-18) gives the closed-loop
expression
Y G% fY 1 fG% D
(12-22)

Chapter 12

sp

Thus, the closed-loop transfer function for set-point changes is

Y
G%
f
Ysp

(12-23)

Selection of c
The choice of design parameter c is a key decision in both the
DS and IMC design methods.
In general, increasing c produces a more conservative
controller because Kc decreases while I increases.
21

Chapter 12

Several IMC guidelines for c have been published for the

model in Eq. 12-10:
1.

c / > 0.8 and c 0.1 (Rivera et al., 1986)

2.

(Chien and Fruehauf, 1990)

3.

(Skogestad, 2003)

Controller Tuning Relations

In the last section, we have seen that model-based design
methods such as DS and IMC produce PI or PID controllers for
certain classes of process models.

IMC Tuning Relations

The IMC method can be used to derive PID controller settings
for a variety of transfer function models.
22

Chapter 12

Table 12.1 IMC-Based PID Controller Settings for Gc(s)

(Chien and Fruehauf, 1990). See the text for the rest of this
table.

23

24

Chapter 12

Tuning for Lag-Dominant Models

Chapter 12

First- or second-order models with relatively small time delays

/ = 1 are referred to as lag-dominant models.
The IMC and DS methods provide satisfactory set-point
responses, but very slow disturbance responses, because the
value of I is very large.
Fortunately, this problem can be solved in three different ways.
Method 1: Integrator Approximation
s
s
Ke
K
*
e
Approximate G%( s )
by G%
( s)
s 1
s
where K * @K / .

Then can use the IMC tuning rules (Rule M or N)

to specify the controller settings.
25

Method 2. Limit the Value of I

Chapter 12

For lag-dominant models, the standard IMC controllers for firstorder and second-order models provide sluggish disturbance
responses because I is very large.
For example, controller G in Table 12.1 has I where is
very large.
As a remedy, Skogestad (2003) has proposed limiting the value
of I :
I min 1 , 4 c

(12-34)

where1 is the largest time constant (if there are two).

Method 3. Design the Controller for Disturbances, Rather
Set-point Changes
The desired CLTF is expressed in terms of (Y/D)des, rather than (Y/Ysp)des
Reference: Chen & Seborg (2002)

26

Example 12.4
Consider a lag-dominant model with / 0.01:

Chapter 12

G% s

100 s
e
100 s 1

Design four PI controllers:

a) IMC c 1
b) IMC c 2 based on the integrator approximation
c) IMC c 1 with Skogestads modification (Eq. 12-34)
d) Direct Synthesis method for disturbance rejection (Chen and
Seborg, 2002): The controller settings are Kc = 0.551 and
I 4.91.

27

Evaluate the four controllers by comparing their performance for

unit step changes in both set point and disturbance. Assume that
the model is perfect and that Gd(s) = G(s).

Chapter 12

Solution
The PI controller settings are:
Controller

Kc

(a) IMC

0.5

100

(b) Integrator approximation

0.556

(c) Skogestad

0.5

(d) DS-d

0.551

4.91

28

Chapter 12

Figure 12.8. Comparison

of set-point responses
(top) and disturbance
responses (bottom) for
Example 12.4. The
responses for the Chen
and Seborg and integrator
approximation methods
are essentially identical.

29

Chapter 12

Tuning Relations Based on Integral

Error Criteria
Controller tuning relations have been developed that optimize
the closed-loop response for a simple process model and a
specified disturbance or set-point change.
The optimum settings minimize an integral error criterion.
Three popular integral error criteria are:
1. Integral of the absolute value of the error (IAE)

IAE e t dt

(12-35)

where the error signal e(t) is the difference between the set
point and the measurement.
30

Chapter 12
12a

Figure 12.9. Graphical

interpretation of IAE.
The shaded area is the
IAE value.

31

2. Integral of the squared error (ISE)

ISE e t

dt

(12-36)

Chapter 12

3. Integral of the time-weighted absolute error (ITAE)

ITAE t e t dt

(12-37)

See text for ITAE controller tuning relations.

Comparison of Controller Design and

Tuning Relations
Although the design and tuning relations of the previous sections
are based on different performance criteria, several general
conclusions can be drawn:
32

1. The controller gain Kc should be inversely proportional to the

product of the other gains in the feedback loop (i.e., Kc 1/K
where K = KvKpKm).
2. K should decrease as / , the ratio of the time delay to the

Chapter 12

dominant time constant, increases. In general, the quality of

/
control decreases as
increases owing to longer settling
times and larger maximum deviations from the set point.
I
D
/
3. Both
and
should increase as /increases.
For many

D
I
controller tuning relations, the
ratio,
, is between 0.1 and
D / I
0.3. As a rule of thumb, use
= 0.25 as a first guess.
4. When integral control action is added to a proportional-only
controller, Kc should be reduced. The further addition of
derivative action allows Kc to be increased to a value greater
than that for proportional-only control.
33

Chapter 12

Controllers With Two Degrees

of Freedom
The specification of controller settings for a standard PID
controller typically requires a tradeoff between set-point
tracking and disturbance rejection.
These strategies are referred to as controllers with twodegrees-of-freedom.
The first strategy is very simple. Set-point changes are
introduced gradually rather than as abrupt step changes.
For example, the set point can be ramped as shown in Fig.
12.10 or filtered by passing it through a first-order transfer
function,
*
Ysp
1

(12-38)
Ysp f s 1
34

*
where Ysp
denotes the filtered set point that is used in the control
calculations.

Chapter 12

The filter time constant, f determines how quickly the filtered

set point will attain the new value, as shown in Fig. 12.10.

Figure 12.10 Implementation of set-point changes.

35

 A second strategy for independently adjusting the set-point

response is based on a simple modification of the PID control
law in Chapter 8,
t

dym
1
*
*
p t p K c e t e t dt D
(8-7)

I 0
dt

where ym is the measured value of y and e is the error signal.

e @ysp y. m

Chapter 12

The control law modification consists of multiplying the set

point in the proportional term by a set-point weighting factor, :
p t p K c ysp t ym t
1 t * *
dym
Kc
e t dt D

dt
I 0

(12-39)

The set-point weighting factor is bounded, 0 < < 1, and serves

as a convenient tuning factor.

36

Chapter 12
Figure 12.11 Influence of set-point weighting on closed-loop
responses for Example 12.6.
37

On-Line Controller Tuning

Chapter 12

1. Controller tuning inevitably involves a tradeoff between

performance and robustness.
2. Controller settings do not have to be precisely determined. In
general, a small change in a controller setting from its best
value (for example, 10%) has little effect on closed-loop
responses.
3. For most plants, it is not feasible to manually tune each
controller. Tuning is usually done by a control specialist
(engineer or technician) or by a plant operator. Because each
person is typically responsible for 300 to 1000 control loops, it
is not feasible to tune every controller.
4. Diagnostic techniques for monitoring control system
performance are available.
38

Chapter 12

Continuous Cycling Method

Over 60 years ago, Ziegler and Nichols (1942) published a
classic paper that introduced the continuous cycling method for
controller tuning. It is based on the following trial-and-error
procedure:
Step 1. After the process has reached steady state (at least
approximately), eliminate the integral and derivative control
action by setting D to zero and I to the largest possible value.
Step 2. Set Kc equal to a small value (e.g., 0.5) and place the
controller in the automatic mode.
Step 3. Introduce a small, momentary set-point change so that the
controlled variable moves away from the set point. Gradually
increase Kc in small increments until continuous cycling occurs.
The term continuous cycling refers to a sustained oscillation with
a constant amplitude. The numerical value of Kc that produces

39

Chapter 12

continuous cycling (for proportional-only control) is called the

ultimate gain, Kcu. The period of the corresponding sustained
oscillation is referred to as the ultimate period, Pu.
Step 4. Calculate the PID controller settings using the ZieglerNichols (Z-N) tuning relations in Table 12.6.
Step 5. Evaluate the Z-N controller settings by introducing a
small set-point change and observing the closed-loop response.
Fine-tune the settings, if necessary.
The continuous cycling method, or a modified version of it, is
frequently recommended by control system vendors. Even so, the
continuous cycling method has several major disadvantages:
1. It can be quite time-consuming if several trials are required and
the process dynamics are slow. The long experimental tests
may result in reduced production or poor product quality.
40

Chapter 12

Pu

Figure 12.12 Experimental determination of the ultimate gain

Kcu.
41

42

Chapter 12

Chapter 12

2. In many applications, continuous cycling is objectionable

because the process is pushed to the stability limits.
3. This tuning procedure is not applicable to integrating or
open-loop unstable processes because their control loops
typically are unstable at both high and low values of Kc,
while being stable for intermediate values.
4. For first-order and second-order models without time delays,
the ultimate gain does not exist because the closed-loop
system is stable for all values of Kc, providing that its sign is
correct. However, in practice, it is unusual for a control loop
not to have an ultimate gain.

43

Relay Auto-Tuning

Chapter 12

strm and Hgglund (1984) have developed an attractive

alternative to the continuous cycling method.
In the relay auto-tuning method, a simple experimental test is
used to determine Kcu and Pu.
For this test, the feedback controller is temporarily replaced by
an on-off controller (or relay).
After the control loop is closed, the controlled variable exhibits
a sustained oscillation that is characteristic of on-off control
(cf. Section 8.4). The operation of the relay auto-tuner includes
a dead band as shown in Fig. 12.14.
The dead band is used to avoid frequent switching caused by
measurement noise.
44

Chapter 12

Figure 12.14 Auto-tuning using a relay controller.

45

 The relay auto-tuning method has several important

advantages compared to the continuous cycling method:

Chapter 12

1. Only a single experiment test is required instead of a

trial-and-error procedure.
2. The amplitude of the process output a can be restricted
by adjusting relay amplitude d.
3. The process is not forced to a stability limit.
4. The experimental test is easily automated using
commercial products.

46

Step Test Method

Chapter 12

In their classic paper, Ziegler and Nichols (1942) proposed a

second on-line tuning technique based on a single step test.
The experimental procedure is quite simple.
After the process has reached steady state (at least
approximately), the controller is placed in the manual mode.
Then a small step change in the controller output (e.g., 3 to
5%) is introduced.
The controller settings are based on the process reaction curve
(Section 7.2), the open-loop step response.
Consequently, this on-line tuning technique is referred to as the
step test method or the process reaction curve method.

47

Chapter 12
Figure 12.15 Typical process reaction curves: (a) non-selfregulating process, (b) self-regulating process.
48

Chapter 12

An appropriate transfer function model can be obtained from the

step response by using the parameter estimation methods of
Chapter 7.
The chief advantage of the step test method is that only a single
experimental test is necessary. But the method does have four
disadvantages:
1. The experimental test is performed under open-loop conditions.
Thus, if a significant disturbance occurs during the test, no
corrective action is taken. Consequently, the process can be
upset, and the test results may be misleading.
2. For a nonlinear process, the test results can be sensitive to the
magnitude and direction of the step change. If the magnitude of
the step change is too large, process nonlinearities can
influence the result. But if the step magnitude is too small, the
step response may be difficult to distinguish from the usual
fluctuations due to noise and disturbances. The direction of the
step change (positive or negative) should be chosen so that 49

the controlled variable will not violate a constraint.

Chapter 12

3. The method is not applicable to open-loop unstable processes.

4. For analog controllers, the method tends to be sensitive to
controller calibration errors. By contrast, the continuous
cycling method is less sensitive to calibration errors in Kc
because it is adjusted during the experimental test.

Example 12.8
Consider the feedback control system for the stirred-tank blending
process shown in Fig. 11.1 and the following step test. The
controller was placed in manual, and then its output was suddenly
changed from 30% to 43%. The resulting process reaction curve is
shown in Fig. 12.16. Thus, after the step change occurred at t = 0,
the measured exit composition changed from 35% to 55%
(expressed as a percentage of the measurement span), which is
equivalent to the mole fraction changing from 0.10 to 0.30.
Determine an appropriate process model for G @GIP Gv G p Gm .
50

Chapter 12
Figure 11.1 Composition control system for a stirred-tank
blending process.
51

Chapter 12
Figure 12.16 Process reaction curve for Example 12.8.
52

Chapter 12

Figure 12.17 Block diagram for Example 12.8.

53

Chapter 12

Solution
A block diagram for the closed-loop system is shown in Fig.
12.17. This block diagram is similar to Fig. 11.7, but the feedback
loop has been broken between the controller and the current-topressure (I/P) transducer. A first-order-plus-time-delay model can
be developed from the process reaction curve in Fig. 12.16 using
the graphical method of Section 7.2. The tangent line through the
inflection point intersects the horizontal lines for the initial and
final composition values at 1.07 min and 7.00 min, respectively.
The slope of the line is

55 35%
S
3.37% / min

7.00 1.07 min

and the normalized slope is
S 3.37% / min
R

0.259 min 1
p 43% 30%

54

Chapter 12

The model parameters can be calculated as

xm 55% 35%
K

1.54 dimensionless
p 43% 30%
1.07 min
7.00 1.07 min 5.93 min
The apparent time delay of 1.07 min is subtracted from the
intercept value of 7.00 min for the calculation.
The resulting empirical process model can be expressed as
X m s
1.54e 1.07 s
G s
P s
5.93s 1
Example 12.5 in Section 12.3 provided a comparison of PI
controller settings for this model that were calculated using
different tuning relations.
55

Guidelines For Common Control Loops

(see text)
Chapter 12

Troubleshooting Control Loops

If a control loop is not performing satisfactorily, then
troubleshooting is necessary to identify the source of the
problem.
Based on experience in the chemical industry, he has observed
that a control loop that once operated satisfactorily can become
either unstable or excessively sluggish for a variety of reasons
that include:
a. Changing process conditions, usually changes in
throughput rate.
b. Sticking control valve stem.
56

c. Plugged line in a pressure or differential pressure

transmitter.

Chapter 12

d. Fouled heat exchangers, especially reboilers for

distillation columns.
e. Cavitating pumps (usually caused by a suction pressure
that is too low).
The starting point for troubleshooting is to obtain enough
background information to clearly define the problem. Many
questions need to be answered:
1. What is the process being controlled?
2. What is the controlled variable?
3. What are the control objectives?
4. Are closed-loop response data available?
5. Is the controller in the manual or automatic mode? Is it
reverse or direct acting?

57

6. If the process is cycling, what is the cycling frequency?

7. What control algorithm is used? What are the controller
settings?

Chapter 12

8. Is the process open-loop stable?

9. What additional documentation is available, such as
control loop summary sheets, piping and instrumentation
diagrams, etc.?

58