Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Pid Post

Download as pdf or txt
Download as pdf or txt
You are on page 1of 33

2.

PID control

2 PID CONTROL
This paper investigates the PID control of the Single-Input Single-Output (SISO)
processes. The basic configuration of the control system is shown in the block diagram below
(Fig. 2.1). The block diagram gives the notation used for the signals in the control system.
The error signal e(t) is formed as the difference between the reference signal r(t) and the
output signal y(t). The PID controller acts on the error signal to form control signal u(t). Some
PID controllers with modified structure also use output signal to form control signal. Disturbance
z(t) can be introduced into the process at different points. Furthermore, various noise sources
(e.g. process noise, measurement noise) can be present in the control system, but it is adequate to
model noise influence on the control system with the signal n(t), which is added to the output of
the process to form the output signal.

error
signal
e(t)

reference
signal
r(t)
+
-

PID
Controller

control
signal
u(t)

disturbance
z(t)

noise
n(t)
+

Process

ouput
signal
y(t)

Fig. 2.1. Basic control system configuration.


The block marked Process in Figure 2.1 includes all the elements of the control system
which are considered as the parts of the process: actuator, plant, and sensor. The mathematical
model of the process can be very complex, with complicated static and dynamic description. The
identification of a complex model requires a lot of engineering effort. Since the performance
obtained from control system with the PID controller is limited, many PID controller tuning
methods use simple models which have similar complexity as the PID controller. These models
require simple identification experiments and capture dominant dynamic properties. Usual
representation of these models is low-order (first or second order) transfer function in Laplace
domain. On the other hand, some tuning methods for PID controllers were developed for more
complex process models (e.g. higher order models, models with non-linear characteristics)
because of successful and widespread use of PID controllers in industry (Persson, 1992). A
transfer function modeling the process is generally represented as follows:
1 + b1s + b2 s + b2 s2 +...+ bm s m
GP ( s ) = K k
e Tt s , with m n,
2
n
s 1 + a1s + a2 s + a2 s +...+ an s

(2-1)

where Tt is dead time, m and n are degrees of complex variable polynomials, and k is the number
of integrators present in the process. The type of the process model is determined by the

2. PID control

exponent k, so that a process without an integrator (k=0) is called type 0 process, a process
with an integrator (k=1) is called type 1 process, and so on.
Two common process models are used in this study. The first model is the First Order
with Dead Time (FODT) model, often used for the description of chemical processes, given by
the transfer function:
GP ( s ) = G1 ( s ) =

K1e Tt 1s
.
1 + T1s

(2-2)

The second model is frequently employed to describe electromechanical processes. It consists of


an integration and a first order lag:
GP ( s) = G2 (s ) =

K2
.
s(1 + T2 s )

(2-3)

A PID controller consists of the three terms: proportional (P), integral (I), and derivative
(D). Its behavior can be roughly interpreted as the sum of the three term actions: the P term gives
a rapid control response and a possible steady state error; the I term eliminates the steady state
error; and the D term improves the behavior of the control system during transients. This
description of the term actions matches the actual behavior of the PID control system, when it is
used for some processes and for some excitation signals (for example, type 0 process and step
set-point change).
A PID-type controller can be implemented variously. The implementation choice
depends on the structure of the process, design specifications, and PID controllers features. The
following subsections describe how controllers apply the PID control law through the review of
different PID controller forms and implementation aspects. The section proceeds with an outline
of different tuning rules and explains the usage of PI controllers in dead-time compensating
controllers.

2.1 Forms of the PID controller


Different forms of PID controller reflect the development of the PID algorithm in
different technologies and its use in diverse control systems. Besides, some PID forms ensure
better performance and behavior of the control system than others. The textbook version of the
PID control law in the time domain is:

z
t

u( t ) = K Pe( t ) + K I e( )d + K D
0

de( t )
1 t
de( t )
= K P ( e( t ) +
e( )d + TD
),
dt
TI 0
dt

(2-4)

where KP is proportional gain, KI is the gain of the integral term, KD is the gain of the derivative
term, TI is the integral time constant, and TD is the derivative time constant.
The Laplace transformation of equation (2-4) gives the transfer function of the PID
controller:

2. PID control

5
GR ( s ) =

1
1 + TI s + TDTI s2
U ( s)
= K P (1 +
+ TD s ) = K P
.
E( s)
TI s
s

(2-5)

Since the numerator of the PID controller transfer function in (2-5) has a higher degree than the
denominator, the transfer function is not causal and as such can not be realized. The form (2-5)
of the PID controller is modified in such way as to make the controller. It is achieved through the
addition of a lag to the derivative term:
1
U ( s)
TD s
= K P (1 +
+
),
(2-6)
E( s)
TI s 1 + TD s
N
where TD/N is the time constant of the added lag. Divisor N in (2-6) determines the gain KHF of
the PID controller in the high frequency range:
GR ( s ) =

K HF = lim GR ( j ) = K P (1 + N ) .

(2-7)

The gain KHF must be limited because measurement noise signal n(t) often contains high
frequency components and its amplification should be limited. Usually, the divisor N is chosen
in the range 310 (Hang et al., 1991).
The form of the PID controller defined by (2-5) is called the parallel form, because
proportional, integral, and derivative term act simultaneously on the error signal and the control
signal is the sum of the term actions. It is also called non-interactive form of the PID controller
(see Figure 2.2-a.).

P
D
E(s)

+
I

U(s)
+

E(s)
+

U(s)

D
a) Parallel form;

b) Series form;

Fig. 2.2. Forms of the PID controller.


Another form of the PID controller is the series or interacting form (Fig. 2.2.b.) with transfer
function:
GRS ( s ) = K PS (1 +

1
)(1 + sTDS ) ,
sTIS

(2-8)

This form of the PID controller has a simple representation in the frequency domain, since all
roots and zeros of GRS(s) are real and correspond to the inverses of the break frequencies. This
form of the PID controller is also referred to as a classical form of the PID controller, since it

2. PID control

conforms to the structure of PID controllers when these were implemented in the pneumatic
technique.
Based on equations (2-6) and (2-8), the relations for converting the parameters between
the parallel and the series form of the PID controller are:
K PS =

KP
(1 + 1 4TD / TI ),
2

(2-9)

TIS =

TI
(1 + 1 4TD / TI ),
2

(2-10)

TI
(1 1 4TD / TI ),
(2-11)
2
and can be used only if TI 4TD., i.e. when poles and zeros of the parallel form are real.
TDS =

As output of the above forms of the PID controller is the total value of the control signal
u(t), they are called positional PID algorithms. Some actuators such as a motor may use the
increment or derivative of the control signal as an input signal, because they have built-in
integral action. PID controllers with such an output are termed velocity (continuos version) or
incremental (discrete version) PID controllers. An advantage of an incremental version of the
PID algorithm is that it allows straightforward implementation of the algorithm extensions like
anti-windup scheme and bumpless parameter switch (Isermann, 1989) which are described in the
following subsection.
PID controller
FP
R(s)

+
KP

FI

+
-

FD

+
-

KP
TIs

+
+

U(s)
+

KPTDs
1+sTD/N
Y(s)

Fig. 2.3. Two-degrees of freedom PID controller


Standard PID controllers act on the error signal e(t) and give the control signal u(t) as the
output. Such configuration of the controller uses the same parameters as in responding to setpoint change (tracking) and to load disturbance (regulating). The two functions of the control
system often impose contradictory demands on the value of the controller parameters. The
contradiction is resolved by a trade-off in the controller's design. The trade-off can be recognized

2. PID control

in the responses of the control system to set-point change and to load disturbance, which can not
have the same quality. In order to avoid this trade-off a modification of the PID controller
structure was devised (Fig. 2.3.). Signal channels for reference signal and for measurement
signal are separated, and a set of weights (FP, FI, FD) in the channel of the reference signal is
introduced. Weights in the channel of the reference signal allow design, which arbitrarily assigns
zeros of the closed-loop transfer function, and therefore define dynamic behavior of the closedloop system to set-point change.
The PID controller with set-point weighting is tuned in two steps:
1. Parameters are tuned for good regulation;
2. Weights F are adjusted in order to set zeros of the closed-loop transfer function and
thus to improve the tracking behavior of the control system.
Controllers that allow such separation of the design for regulating and for tracking are
called two-degree-of-freedom controllers. Eitelberg (1987) and Hippe et al. (1987) gave some
recommendations for the appropriate choice of weights and proposed the use of filters instead of
constant weights.
The introduction of arbitrary weights in the reference channel of the PID controller gives
design more freedom and renders it more complicated. In some PID controller implementations
(Stoji and Petrovi, 1986; Hang at el., 1989) weights are set to FD=0 and FP=0 in order to avoid
derivative and proportional bumps, which are present in response to step set-point change.
Furthermore, the weight of integral term is set to FI=1, which ensures that the steady state error
for unit step set-point change equals zero.
The standard form of the PID controller, similar to the structure depicted in Figure 2.3,
and recommended by Instrument Society of America, is given below (strom and Hgglund,
1995):
U ( s) = K R [( FP R(s ) Y ( s)) +

1
sTD
( R( s) Y ( s)) +
( FD R(s) Y (s ))].
sTI
1 + sTD / N

(2-12)

Frequently, only a part of a PID controller is used. strom and Hgglund (1995) have
noted that most control loops are of the PI type. As a rule, the PI controller is used for processes
of the first order, or for processes not requiring tight control. A PD controller can be used for
processes which contain integrators, and which do not have constant load disturbances, since PD
controller can not compensate for it. The application of a P controller is limited to simple control
tasks.
Today, almost all control strategies are implemented as digital algorithms in various
devices such as Programmable Logic Controllers (PLCs), Digital Signal Processors (DSPs), and
in other microprocessor-based equipment. To become applicable in such equipment, the PID
control algorithm has to be discretized. Using Euler integration method rectangular integration,
the discrete version of the positional algorithm (2-4) is calculated as:

LM
N

u( k ) = K P e( k ) +

OP
Q

T k 1
T
e( i ) + D ( e( k ) e( k 1)) ,
TI i= 0
T

(2-13)

where k denotes discrete time instant, and T is the sampling time. Recursive equation describing
the incremental version of the PID algorithm is obtained when the equation (2-13) for the time
instant k-1 is subtracted from the same equation for the time instant k:

2. PID control

8
u(k ) = u(k ) u( k 1) = q0 e(k ) + q1e(k 1) + q2 e(k 2 ) ,

where:

(2-14)

q0=KP (1+TD/T),
q1=KP (1+ 2 TD/T T/TI ) and

q2=KP TD / T.
Other relations for parameters qi in (2-14) are obtained if a different integration method (e.g.
trapezoidal method) is used. These discrete approximations of the continuos PID controller are
valid only if the sampling time T is sufficiently short in comparison to the time constants of the
controller. Otherwise, when sampling time T is not much shorter than the time constants of the
controller, connection with continuos PID controllers is dropped and Z-transform form of the
PID controller is used (Isermann, 1989). Discrete controller of the second order with an
integrator has a transfer function in the Z-domain:
q0 + q1 z 1 + q2 z 2
.
(2-15)
1 z 1
Polynomial coefficients qi in (2-15) have to satisfy the following relations (Isermann, 1989):
GC ( z ) =

q0 > 0,
q1<-q0,
-(q0+q1)<q2<q0,
so that the obtained digital controller has a dynamic behavior of the continuos PID controller.

2.2 Practical issues in the application of PID control


All controllers are designed to work with processes which have some physical constraint:
valves have a limited operating range (0%-100%), pumps have limited power, motors have a
maximum moment, and so on. These limitations can be regarded as non-linearities in the process
and have to be considered in the application of the controller. Since many of these limitations
appear at the input of the actuator, they are referred to as input limitations and are modeled with
a non-linear element having a saturation characteristic, as shown in Figure 2.4. It is assumed that
the disturbance z(t) adds to the input signal of the process and adequately models actual
disturbances in the control system. Beside input magnitude limitations, actuators often have
defined rate limitations, or maximum rates at which the control signal u(t) can be changed.
When the controller output signal u0(t) exceeds the upper limit umax, or when it falls
below the lower limit umin of the operating range, its value changes in the input limitation
element so that the controller output signal u0(t) and the process input u(t) do not coincide.
Shouldnt the two coincide, the integrator in a contoller with integral action would produce an
inaccurate and highly excessive value which would cause osillation and slowing down of the
transient response. In other words, the effect would be a large overshoot and a long settling time.

2. PID control

This behavior is called the integrator windup. Moreover, the feedback loop during the windup
behaves as if it were broken. There are several anti-windup algorithms to avoid adverse effects
of the integrator windup on the control system performance. This paper describes the most
frequently used anti-windup algorithms. The reader may further consult comparative studies of
these and other algorithms in Bohn and Atherton (1995) and Vrani (1997).

r(t)
+

e(t)
Controller

controller
output
u0(t)

process z(t)
input
umax
u(t) +
Process
+

umin

n(t)
+

y(t)

input limitation

Fig. 2.4. Input limitation as a part of the control system.


Figure 2.5 shows the structure of the linear feedback anti-windup algorithm. It relies on
the assumption that it is possible to measure both the controller output signal u0(t) and the
process input signal u(t). If the measurement of the process input signal is not possible, the
simulation of the saturation element in the controller can be used. A new signal eaw(t) is added as
an additional input to the integrator of the controller. It is active when there is difference between
the controller output u0(t) and the process input u(t). It acts in direction opposite to the windup
effect. The rate of anti-windup action is defined with the constant TAW which can be explained as
a time constant of this action. strom and Hgglund (1995) calculated its value as follows:
TAW = TI TD .
-Y(s)

KPTDs
1+sTD/N
+

E(s)

KP
KP
TI

+
+

U0(s)
+

1
s
Eaw(s)

U(s)

1
TAW

Fig. 2.5. Structure of the linear feedback anti-windup algorithm.


Another anti-windup algorithm, suitable for discrete implementation, is the conditional
integration algorithm. It allows integration of the error signal e(t) in the integrator element

2. PID control

10

provided that some conditions imposed on the signals present in the control system are met.
Otherwise, the integration is not permitted. The condition that stops the integration can be
expressed as (control signal u0(t) is in the saturation) AND (the sign of the integral increment
moves state of integrator deeper into saturation). Furthermore, it is advisable to add a small
hysteresis, so that control signal does not oscillate around the limit value.
Similar deterioration of the control system performance happens when the source of the
control signal u(t) is changed, for example, when the controller is substituted with another, or
when it is switched from manual to automatic mode. The switch between two different modes of
control system operation is called plant-input substitution (Peng et al., 1996). A bump in the
control signal u(t) reflects such plant-input substition unless the switching controllers are
properly prepared. The effect can be avoided by using bumpless transfer techniques. These
techniques calculate the states of the substitution controller before the switch happens, so that
bump in the control signal does not occur. Peng et al. (1996) offer a good survey of bumpless
transfer techniques
Kothare et al. (1994) have introduced a unified theoretical framework for the study of
anti-windup and bumpless transfer (AWBT). It is based on the approach which, as a first step,
designs the linear (PID) controller ignoring non-linear input characteristics of some elements in
the control loop and then adds AWBT compensation to minimize negative effects of these
characteristics on the control loop performance. They have shown that their theory applies in the
analysis of all anti-windup techniques and bumpless transfer, and that two matrices are enough to
parametrize these characteristics.
An important issue in the implementation of discrete control algorithms is the choice of
sampling time. That choice depends on the control-loop dynamics and should follow the
recommendation given in the Shannons theorem. Since there are many signals and elements in
the control loop with different dynamic properties, it is not always clear how to choose the
sampling time. For the discrete PID controller, Isermann (1989) relates the sampling time T to
the settling time T95% of the process (time required for the response to reach 95% of its final
value):
T95%
5 15 .
(2-16)
T
Some rules of thumb have been established for relating sampling time T to the parameters of PI
and PID controllers (strom and Wittenmark, 1990):
PI controller:
T
01
. 0.3 ;
TI

(2-17)

TN
0.2 0.6 ,
TD

(2-18)

PID controller:

2. PID control

11

where N is divisor constant from equation (2-6). Previous relations should serve as guidelines
and, if necessary, should be adjusted for the particular use.
Additionally, the discrete implementation of the PID controller raises several other issues
which have to be addressed:
Effects of finite word length;
Signal quantization effects;
Signal conditioning and prefiltering problems.
Solutions and trade-offs concerning these issues have been addressed in many textbooks (e.g.
strom and Wittenmark, 1990; Isermann, 1989) and in specialized literature (e.g. strom and
Steingrimsson, 1991).

2.3 Tuning methods for PID controllers


Controller tuning methods provide the controller parameters in the form of formulae or
algorithms. They ensure that the obtained control system would be stable and would meet given
objectives. These methods require certain knowledge about the controlled process. This
knowledge, which depends on the applied method, usually translates into a transfer function. The
objectives which should be achieved by the application of the control system are associated with
the following control system features (Persson, 1992):
Regulating performance;
Tracking performance;
Robustness;
Noise attenuation.
Often, the desired objectives put contradictory demands on the values of the controller
parameters, so that various trade-offs have to be made. The objectives can be stated in many
ways such as through:
Specifications within the time domain;
Specifications within the frequency domain;
Robustness specifications;
Other specifications.
The specifications within the time domain give some values related to the shape of
control system signals in the time domain. Figure 2.6 shows a typical output signal y(t), a
response to set-point change. Specification values within the time domain are marked on it:
overshoot m, undershoot u, rise time tr, time of first maximum tm, settling time t, and steady
state error ess. Similar specifications within the time domain are used to describe characteristics

2. PID control

12

of the control system response to load disturbance: peak perturbation dm and disturbance
settling time td.
1.2
y
1
0.8

tm

u
ess

t
tr

0.6
0.4
0.2
0

10

t(s)

15

Fig. 2.6. Specifications in time domain.


The specifications within the frequency domain define some values related to the
frequency characteristics of transfer functions of various elements in the control system.
Bandwidth of a closed-loop control system with transfer function G(s) is the lowest frequency b
for which below relation holds:
G ( j b )
1
=
G( 0)
2

(2-19)

The gain margin AM of the control system, described with the open-loop transfer function GO(s),
is defined as the inverse of the open-loop gain at the phase crossover frequency :
Ar =

1
,
GO ( j )

(2-20)

where the frequency is defined as the lowest frequency with


arg GO ( j ) = .

(2-21)

The phase crossover frequency is also called the ultimate frequency of the control system.
The phase margin of the control system is defined as the phase of the open-loop transfer
function GO(s) at the gain crossover frequency c:

= arg GO ( j c ) + ,

(2-22)

where the frequency c is defined as the lowest frequency with


GO ( j c ) = 1.

(2-23)

Maximum sensitivity Ms of the control system, also called modulus margin, is defined as:
M s = max

1
,
1 + GO ( j )

(2-24)

2. PID control

13

and can be interpreted as the inverse of the shortest distance between the critical point C(-1, i0)
in the Nyquist plane and the Nyquist curve. The above definition (2-24) for the maximum
sensitivity Ms makes it possible to relate Ms to gain and phase margin of the control system
(Persson, 1992):
Ar

Ms
,
Ms 1

2 arcsin

(2-25)

FG 1 IJ .
H 2M K

(2-26)

Figure 2.7. gives an example with the Nyquist curve of a process and specifications within the
frequency domain.
Robustness specifications define allowed deviation of the process parameters from
nominal values. The control system should retain designed stability and performance in the range
of these deviations. The parameter deviations from nominal values can be defined as
multiplicative or additive parameter uncertainty characteristics, expressed in the frequency
domain (Morari and Zafiriou, 1989). Besides, the robustness of the control system can be
specified in terms of gain margin Ar, phase margin , and as maximum allowed sensitivity Ms of
the control system.
Other specifications which define a control system requirements include the description
of the process constrains and other implementation-related specifications.
Im
0.5

0 C

1
Ar

1
Ms

-0.5

-1
Nyquist curve
-1.5

-1

-0.5

0.5

Re

Fig. 2.7. Specifications in frequency domain.


Some controller tuning methods include recommendations of a suitable controller
structure and its parameters. This paper describes tuning methods for the fixed structure of PID
controllers. P, PI, and PD controllers are considered as special cases of PID controller. The
tuning methods for PID controllers can be grouped according to their nature and usage, as
follows:

2. PID control

14

Heuristic methods evolved from practical experience in PID controller tuning;


Frequency methods employ frequency characteristics of the controlled process to tune
PID controller parameters;
Analytical methods calculate PID controller parameters from analytical or algebraic
relations that define control system by direct calculation;
Loop-shaping methods seek to shape the open-loop transfer function of the control
system into a desirable from;
Optimization methods obtain PID controller parameters from different optimization
algorithms;
Methods in which PID controller represents a restriction of possible controller structure
(e.g. PID controller tuning in the framework of Internal Model Control);
Methods for tuning a PID controller which functions as a part of an advanced control
strategy (e.g. usage of PI controller in dead-time compensating controllers).
The above groups do not sharply distinguish and some methods may belong to more than
one group. An important criterion in the evaluation of the presented tuning methods is the
suitability of a particular method for the on-line usage. This especially refers to the possibility to
use a particular method for autotuning.

2.3.1

Ziegler-Nichols tuning rules

Ziegler and Nichols have introduced a useful methodology for controller tuning. It
consists of a simple experiment with a controlled process and extracts some of its features. Once
the experiment is completed, the method provides tables by which it is possible to calculate the
controller parameters. The tuning tables were developed through numerous experiments which
involved different processes. The goal of the design was to find a controller which gives the
quarter amplitude damping (QAD) ratio of the control systems in response to load disturbance.
QAD ratio is achieved when the ratio of the first overshoot and the first undershoot of the control
system response equals (strom and Hgglund, 1995). This design specification arises from
empirical observations and has been used traditionally, but gives too oscillatory control systems.
Ziegler and Nichols considered P, PI, and PID controllers in their work.
The first experiment consists of measuring apparent dead-time TZN1 and the maximum
slope of the response on the process reaction curve (response to step set-point change) KZN1. The
measurements are shown in Figure 2.8.a, and relations for obtaining controller parameters in
Table 2.1.a.

2. PID control
y(t)

15
y(t)0.15

8
7

0.1

6
0.05

5
4

TZN1

-0.05

KZN1

1
0
0

TU

-0.1

10

-0.15
0

time[s]
a) Experiment with the process
reaction curve;

10

time[s]
b) Experiment with the process on
the stability limit;

Fig. 2.8. Ziegler-Nichols experiments: output signal of the control system.


In the second experiment, the process is controlled with a proportional controller. The
gain of the controller is gradually increased until the control system reaches stable oscillations on
the stability limit. The value of the controller gain KU is called the ultimate gain, and the
oscillation period TU is called the ultimate period. The two values serve as the basis for the
calculation of the controller parameters (see Table 2.1.b). Figure 2.8.b shows a typical process
output y(t) during such an experiment.
Controller
P
PI
PID

TI
3 TZN1
2 TZN1

KP
1/(TZN1 KZN1)
0.9/(TZN1 KZN1)
1.2/(TZN1 KZN1)

TD
TZN1/2

a) Experiment with process reaction


curve (ZN1);

Controller
P
PI
PID

KP
0.5 KU
0.45 KU
0.6 KU

TI
0.85 TU
0.5 TU

TD
0.125 TU

b) Experiment with process on the


stability limit (ZN2);

Tab. 2.1. Ziegler-Nichols relations for calculating controller parameters.


Takahashi has developed similar relations (Table 2.2) to calculate the discrete P, PI, and
PID controller parameters. These relations use the incremental form of the PID controller, as
follows:

LM
N

u( k ) = u( k 1) + K P y( k 1) y ( k ) +

OP
Q

T
T
e( k ) + D 2 y ( k 1) y ( k 2) y ( k ) . (2-27)
TI
T

An additional parameter to the original relations in Table 2.2 is the sampling time T. In
relation to other time constants, the sampling time T has to be small enough for the relations to
produce useful results (Isermann, 1989).

2. PID control

16

Many authors assessed the performance of the control systems with controllers tuned
according to Ziegler-Nichols (ZN) rules (e.g. strom and Hgglund, 1995; Hang et al., 1991;
Thomas, 1991). The comparison of the two ZN methods shows that the second can be regarded
as better, since the first method fails to make it clear how to measure apparent dead time TZN1.
Moreover, for many processes the controller gain obtained with the first method is 25%40%
higher than the gain obtained with the second method, giving more oscillatory response to setpoint change. Hang et al. (1991) showed that the performance of the controllers tuned according
to ZN rules depended strongly on the value of the process dead-time, and that ZN rules often
gave poor damping and excessive overshoot in response to set-point change. ZN tuning rules for
PI controller show worse performance than rules for the PID controller and even produce
unstable control systems for some processes (Thomas, 1991).
Controller
P

KP
1
K ZN 1 (TZN 1 + T )

T/TI
-

TD/T
-

PI

0.9
0135
. T

K ZN 1(TZN 1 + T ) K ZN 1 (TZN 1 + T / 2 )2

0.27T
K P K ZN 1(TZN 1 + T / 2 )2

PID

12
.
0.3T

K ZN 1(TZN 1 + T ) K ZN 1 (TZN 1 + T / 2 )2

0.6T
K P K ZN 1(TZN 1 + T / 2 )2

0.5
K P K ZN 1T

a) Experiment with process reaction curve;


Controller
P
PI
PID

FG
H

KP
0.5 KU

KU 0.45 0.27

FG
H

T
TU

IJ
K

T/TI
-

IJ
K

0.54

KU T
K PTU

TD/T
-

KU T
3KU T
T
12
.
K PTU
40 K PTU
TU
b) Experiment with process on the stability limit;
0.6 KU 1

Tab. 2.2. Relations for controller tuning developed by Takahashi.


Hang et al. (1991) addressed these problems and proposed refinements of the original
ZN2 method using the normalized gain of the process:

= K p KU ,

(2-28)

where Kp is the steady state gain of the process.


Normalized dead time of the process is defined with:
=

Tt
,
Tp

(2-29)

where Tt is the apparent dead time, and Tp is the dominant time constant of the process. For many
processes values (2-28) and (2-29) are inversely proportional, so when the normalized gain of the

2. PID control

17

process is large, the normalized dead time is small. Strong dependence between values (2-28)
and (2-29) on one hand, and overshoot in the response to set-point change of the control system
on the other suggested (Hang et al., 1991):
1 The use of set-point weighting in order to reduce excessive overshoot for processes
with small dead-time (2.25<<15):
FD=0; FI=1;
Weight FP is adjusted according to:
Specification overshoot m=10%:
15
;
15 +
Specification overshoot m=20%:
FP =

36
.
27 + 5
2 Set-point weighting and modification of the relation for integral time constant TI of
the PID controller for processes with large dead-time (1.5<<2.25):
Specification overshoot m=20%:
FP =

FD=0; FI=1; FP =

F
H

I
K

8 4
+1 ;
17 9

2
TI = TU .
9
3 New formulae for PI controller tuning for processes with 1.2<<15:
Specification overshoot m=10%:

F
H

I
K

1
4
5 12 +
+1 .
; TI = TU
6 15 + 14
5
15
Most authors agree that ZN tuning rules can quickly give approximate values of the
controller parameters, but require fine-tuning for better performance of the control system. The
ZN2 experiment led to the idea of relay experiment, which is an identification technique
customarily used in autotuning.
K P = KU

2.3.2 PID tuning based on integral criteria


Methods based on integral criteria for tuning PID controller involve searching for the
minimum of the cost function I in the general form:

I = t n f e(t ) dt ,
0

where e(t) is the error signal.

(2-30)

2. PID control

18

Optimum controller parameters and the minimum of the penalty function I is found when
its partial derivatives, in respect to controller parameters, equal zero. Equations for the
calculations of the PID controller parameters are:
I
I
I
= 0,
=0 ,
= 0.
K P
TI
TD

(2-31)

Generally, the set of equations (2-31) can not be solved analytically but numerically. Usually the
choice of a particular function f and exponent n leads to formation of the following criteria
(2-30):
Integral Error (IE):

f[e(t)]=e(t), n=0;

Integral Absolute Error (IAE):

f[e(t)]=|e(t)|, n=0;

Integral Time multiplied Absolute Error (ITAE):

f[e(t)]=|e(t)|, n=1;

Integral Squared Error (ISE):

f[e(t)]=e(t)2, n=0;

Integral Squared Time Error (ITSE):

f[e(t)]=e(t)2, n=1;

Integral Time square multiplied Squared Error (IT2SE):

f[e(t)]=e(t)2, n=2.

For some choices of the integrand in (2-30), it is possible to give physical interpretation
of the cost integral I (Persson, 1992). IE criterion is suitable for the static processes with nonoscillatory behavior. For example, this criterion is suitable for the control of a process with an
output leading to a storage tank. An example of the appropriate use of IAE criterion is the octane
control in gasoline production (Persson, 1992). Minimization of the total control energy in
control systems leads to the ISE criterion. Time-weighted integrand functions, with n>0 in
definition equation (2-30), given by the ITAE and ITSE criteria do not have physical
significance (Seborg et al., 1989). These criteria penalize errors that persist for long periods of
time. The ITAE criterion can be used for obtaining very conservative controller settings.
It is important to note that the error signal e(t), used for optimization, can be a result of
set-point change or of load disturbance. It is, therefore, possible to obtain two sets of parameters:
one optimized for set-point change and the other for load disturbance.
A useful property of a PID controller is that the parameter KI (2-4) is directly related to
IE for unit step load disturbance through (strom and Hgglund, 1995):

I IE = e(t )dt =
0

1
T
= I .
KI KP

(2-32)

For non-oscillatory processes, IE and IAE criteria have the same values, and for welldamped oscillatory processes IE can be regarded as an approximation of IAE. Criteria for the
above mentioned processes can be minimized through maximization of the parameter KI.

2. PID control

19

Parameters
optimized for

KP

TI

Set-point change

0.361 KU

0.083 ( 1.935 +1) TU

Load disturbance

1.892 + 0.244
KU
3.249 + 2.097

0.706 0.227
TU
0.7229 + 1.2736

Tab. 2.3.a Tuning rules for a PI controller.


Parameters
optimized for

KP

TI

TD

Set-point change

0.509 KU

0.051 ( 3.302 +1) TU

0.125 TU

Load disturbance

4.434 0.966
KU
512
. + 1.734

1.751 0.612
TU
3.776 + 1.388

0.144 TU

Tab. 2.3.b. Tuning rules for a PID controller.


Zhuang and Atherton (1993) performed ITSE optimizations for the FODT model. After
that, they studied relations between controller parameters and the ultimate gain and ultimate
period of the process. They found out that the parameters of PI and PID controllers can be
expressed in terms of the ultimate gain KU, ultimate period TU, and normalized gain of the
process. These tuning rules are summarized in Table 2.3 and are prepared for an application in
the autotuning procedures based on relay experiment.

2.3.3 Cohen-Coon tuning rules


The Cohen-Coon tuning method is based on the FODT model (2-2) with main design
specification for quarter amplitude decay (QAD) ratio in response to load disturbance. The
design objectives (Persson, 1992; strom and Hgglund, 1995) were to maximize the gain and
minimize the steady-state error and QAD for P and PD controllers. The parameters of the PI
controller were obtained through minimization of the IE criteria and demand for QAD response.
The parameters for PID controller were calculated with same objectives as for the PI controller.
The positioning of the additional controller pole was on the negative real axis. It is placed at the
same distance from the origin as the two complex poles of the controller.
Relations for controller parameters in Table 2.4 are given in terms of parameters:
K1Tt1
,
T1

(2-33)

Tt1
,
Tt1 + T1

(2-34)

=
=

which are calculated from the parameters of the FODT model (2-2).

2. PID control

20
Controller
P
PI
PD
PID

KP

F
I
H
K
0.9 F 0.92 I
1+
H
1 K
1.24 F 0.13 I
1+
H
1 K
1.35 F 0.18 I
1+
H
1 K

TI

TD

33
. 3
Tt1
1 12
.

0.27 0.36
Tt1
1 0.87

2.5 2
Tt1
1 0.39

0.37 0.37
Tt1
1 0.81

1
0.35
1+

Tab. 2.4. Cohen-Coon controller tuning rules.


strom and Hgglund (1995) observed that the Cohen-Coon tuning method suffers from
a too small decay ratio, which results in low damping and high sensitivity of the closed-loop
system. The method can be applied for on-line PID controller tuning, if the parameters of the
FODT model are known.

2.3.4 PID tuning based on gain and phase margin specifications


PID tuning methods based on gain and phase margin specifications (GPM methods)
involve solving definition equations for gain and phase margins, given by (2-20)-(2-23).
Generally, these equations are non-linear and complicated for solving. Therefore, usual design
methods based on these specifications are solved numerically or graphically, using Bode
diagrams (Peri, 1998).
Ho et al. (1995a) analyzed the control system consisting of a PI controller and a FODT
process with transfer function (2-2). When transfer functions of these dynamic elements are put
into (2-20)-(2-23), arctan function appears in relations determining gain and phase crossover
frequency:
+ arctan TI - arctan T1 Tt1=0,
Ar K P K1 = TI

K P K1 = cTI

2T12 + 1
,
2TI 2 + 1
c 2T12 + 1
,
c 2TI 2 + 1

= + arctan c TI - arctan c T1 c Tt1.

(2-35)
(2-36)

(2-37)
(2-38)

In order to simplify the procedure of solving these non-linear equations, the following
approximation was introduced:

2. PID control

21

R|x
4
arctan( x ) S
||2 - 4x
T

; x 1
.

(2-39)

; x >1

The approximation and solving equations (2-35)-(2-38) for PI controller parameters gives:
KP =

F
T = G 2
H
I

T1
,
Ar K1

2
4 Tt1 1

T1

(2-40)

I
JK

(2-41)

where is calculated through:

Ar + 12 Ar ( Ar 1)
.
2
( Ar 1)Tt1

(2-42)

The parameters for PID controller are derived when the model of the process is second order
with dead time (SODT) model. Equations (2-40) and (2-41) are used for parameters KPS and TIS
in the series form of the PID controller and the derivative time constant TDS is used to
compensate for the smaller time constant.
It is important to note that it is not possible to achieve arbitrary gain and phase margin
through the design procedure. The achievable specifications depend on the process
characteristics and on the values of valid controller parameters. Reasonable values,
recommended by Ho et al. (1995a) are Ar=4 and =600. The method is valid for the FODT
processes with the Tt1/T1 ratio below 1.
Fung et al. (1998) studied exact solutions for the set of equations (2-35)-(2-38)
describing control loops with a PI controller and a general linear process. They have proposed a
graphical method for finding PI controller parameters. The method checks the solvability of the
GPM method problem and is a good analytical tool.
Approximation (2-39) of arctan function was successfully applied in a self-tuning PID
controller (Ho et al., 1997). This approach to the design of PID controllers is, therefore,
particularly useful in the context of adaptive control and autotuning.

2.3.5 Approximate pole placement method: Lambda tuning


Pole placement is an approach to control design which places all closed-loop poles in the
desired places. Poles can be placed by employing a feedback controller or by using a state
feedback controller which can be accompanied with an observer (strom and Wittenmark,
1990). The methodology has been developed for processes of an arbitrary degree and can
produce high-order controllers and observers. In order to use pole placement methodology for
tuning PID controllers, only few of the closed-loop poles are placed in the desired places. One of
the methods based on this principle is the Lambda tuning method.

2. PID control

22

Lambda tuning method is an approximate pole placement method which was originally
devised by Dahlin and Highham to control processes with time delays (Panagopoulos et al.,
1997). The controller is designed in such fashion as to place a pole of the closed-loop system at
the location s=-. The method was named after the parameter which is the inverse of the
desired closed-loop time constant TCL=1 / .
The method can be applied to obtain parameters of the PI controller for the control of a
process modeled by the FODT model (2-2). The parameters can be calculated with or without
cancellation of the process pole by the controller zero (Panagopoulos et al., 1997). In the
derivation of the tuning rules, a complex exponential which models the process time delay is
approximated with e-sTt1 1-sTt1. Tuning rules are given in Table 2.5.
Parameters

PI controller with
pole-zero cancellation

PI controller without
pole-zero cancellation

KR

1
T1
K1 Tt1 + TCL

1 TT
1 t 1 + 2TCLT1 TCL
2
K1
Tt1 + TCL

TI

T1

TT
1 t 1 + 2TCLT1 TCL
Tt1 + TCL

Tab. 2.5. Lambda tuning rules for PI controller.


Cancellation of process poles by the controller zero results in the presence of
uncontrollable modes in the closed-loop system, which may lead to poor performance if these
modes are excited. This effect is particularly apparent in the control systems with the cancelled
modes slower than the dominant modes (strom and Hgglund, 1995; Hang, 1989). Besides,
pole-zero cancellation applied in the systems with large signal non-linearities in the loop can
produce large transient signals (Clark, 1988).

2.3.6 Approximate pole placement method: Dominant pole design


Dominant-Pole Design (DPD) methods find controller parameters which place the
dominant poles of the closed-loop system in specified locations. In other words, those methods
can be viewed as a translation of the problem of finding controller parameters into the problem
of placing dominant poles in desired locations. The number of dominant poles to be placed
depends on the number of free parameters, that is, on the number of controller parameters. A PI
controller allows placement of two dominant poles, and a PID controller of three dominant poles.
For these controllers, locations of the closed-loop dominant poles are parameterized with
(strom and Hgglund, 1995):
PI controller:
p1,2 = n ( j 1 2 ) , 0< <1;

(2-43)

2. PID control

23

An additional pole location for PID controller:


p3 = k0 n .

(2-44)

Pole locations (2-43) and (2-44) in s-plane are depicted in Figure 2.9, where angle is
determined through =acos ().
Im

p1

n
s - plane

p3
k0n

Re

p2

Fig. 2.9. Locations of dominant poles for PI and PID controllers.


Calculation of the closed-loop poles involves solving characteristic equation of the
control system:
1+GR(s)GP(s) =0.
(2-45)
If the required pole location is p1, and the controller GR(s) is a PI controller, the characteristic
equation of the control system is:

FG
H

1 + KP

IJ
K

1 + TI p1
GP ( p1 ) = 0 ,
TI p1

(2-46)

and it should be solved for the controller parameters KP and TI. To simplify calculation of the
characteristic equation transfer function GP(s), modeling the process, is parameterized with:
GP ( s ) s=

ne

j ( )

= GP ( n e j ( ) ) = a ( n , )e j ( n , )

(2-47)

Parameter functions a(n,) and (n,) can described as frequency characteristics of the
process on the ray with angle in s-plane, where a can be considered as gain characteristic and
can be considered as phase characteristic. By putting parameterizations (2-43) and (2-47) in
the characteristic equation (2-46), and solving it, following relations for PI controller parameters
are obtained:
KP =

sin ( n , ) +
,
a ( n , )sin( )

(2-48)

2. PID control

24
TI =

sin ( n , ) +
.
n sin( ( n , ))

(2-49)

Dominant pole locations of the control system are set to p1 and p2 with suitable choice of n and
, and through calculation of the PI controller parameters according to (2-48) and (2-49). Similar
relations can be derived for parameters of the PID controller.
Direct choice of pole locations pi may lead to poorly damped control systems, which is
avoided through a kind of optimization procedure. The optimization procedure is defined by
specifications in time domain or frequency domain. Generally, value n in DPD methods is used
to achieve specified performance objective, and damping is used to set the required robustness
of the control system. Persson (1992) has proposed four methods, as follows:
Two pole method (2PM) for a PI controller design; n is chosen through maximization of
the parameter KI, which minimizes IE of the control system, and is chosen through
specification of the desired maximum sensitivity Ms. The usual Ms values keep in the
range 1.5-2.2, where values below Ms=1.8 produce non-oscillatory control systems;
Modified PI controller design (MPI) takes into account that maximization of KI may at
times produce a control system with too large an overshoot in response to load
disturbance. This is why the parameter n is increased until the value of the
parameter KI drops to 0.8 KI-max;
Three pole method (3PM) for a PID controller design; parameters n and are chosen as
in the 2PM and the parameter k0 is selected from values ranging between 0.5 and 1;
the method poses moderate computational demands;
Modified controller method (MCM) for a PID controller design uses the 2PM to
calculate the PI controller parameters with maximum sensitivity MS1. After that, the
parameter KD is increased, and n changes to keep KI maximal until the required MS2
is obtained. This method is very computationally demanding and does not work well
for resonant systems.
The position of a pole, which is defined by the D term lag (divisor N in (2-6)), should not
be a part of the optimization procedures, because it might acquire atypical values. The lag should
be considered as a noise filter, and thus, as an implementation issue.
Persson (1992) observed that the maximum sensitivity Ms of the control system, defined
as a design objective, produces controllers with similar responses in the time domain for
different processes. On the other hand, common measures for robustness gain and phase
margin, set as design objectives can produce controllers with very dissimilar responses in the
time domain for different processes (Fung et al., 1998). This makes the maximum sensitivity a
very good choice for a design objective.
DPD methods rely on the knowledge of the process, that is, of the transfer function which
models it, and are off-line in their nature. However, the design can be made in advance for some
models and the results of the design can be used for on-line computation of controller
parameters.

2. PID control

25

2.3.7 Magnitude optimum and symmetric optimum tuning methods


Magnitude optimum (MO) and symmetric optimum (SO) are two loop-shaping tuning
methods extensively employed by the German company Siemens. The first step in the
application of these methods is to determine appropriate transfer function which models the
process. Once the transfer function is determined, the controller is able to shape the open-loop
transfer function in a desired.
MO tuning method was devised with the objective to obtain a control system with a
frequency as close to unity and as flat as possible for the maximum bandwidth (Umland and
Safiuddin, 1990; Deur, 1999). Its mathematical expression states the requirements posed on the
closed-loop transfer function GC(s):
GC(0)=1,

(2-50)

d n GC ( j )
= 0,
lim
=0
d n
for as many n as possible. The desired open-loop transfer function is:

n2
GO1( s ) =
,
s( s + 2 n )

(2-51)

(2-52)

where is the damping of the closed-loop system and n determines the closed-loop dynamics,
that is, the speed of response. For example, the PI controller is employed when it is possible to
approximate the model of the process with the transfer function:
GP ( s ) =

K
,
(1 + T1s )(1 + T2 s )

(2-53)

with T2<T1,. By analyzing (2-50)-(2-53), and by setting =0.707, PI controller parameters are
calculated (strom and Hgglund, 1995)
KP =

TI
,
2 K T2

TI =T1 ,

(2-54)
(2-55)

with n=0.707/T2. The dominant pole is cancelled by the PI controller zero, and the closed-loop
dynamics are determined the smaller time constant T2 of the process.
MO design method optimizes the closed-loop transfer function GC(s) between the
reference and the output signal. It often cancels the process poles by the controller zeros, which
can lead to poor performance of the control system in response to load disturbance (Umland and
Safiuddin, 1990).
Vrani (1997) used a multiple integration of the open-loop step response to identify the
process model and optimized the PID controller parameters according to the MO principle. The
method is termed Modulus Optimum Multiple Integration (MOMI). It is suitable for the
proportional processes with a time delay (type 0 processes) which can be modeled with
following transfer function:

2. PID control

26
GP ( s) = K PR

1 + b1s + b2 s + b2 s 2 +...+ bm s m sL
e .
1 + a1s + a2 s + a2 s 2 +...+ an s n

(2-56)

Integrals, used in the procedure, are defined as:

z FH K y(U )IKd , A = I (),


I (t ) = z b A I ( )gd , A = I ( ),

I1 (t ) =

PR

(2-57)

I3 ( t ) =

z b A I ( )gd , A = I (),...
t

where KPR is the identified gain of the process, y(t) is the response to the applied step, U is the
amplitude of the applied step, and Ai are integrated areas. Parameters of the PID controller are
calculated in terms of the first five integrated areas (A1 - A5) as (Vrani, 1997):
KP =

A3
,
2 A1 A2 A3 K PR TD A12

(2-58)

A3
,
A2 TD A1

(2-59)

A3 A4 A2 A5
.
A3 A3 A1 A5

(2-60)

TI =
TD =

In practice, significant identification errors occur due to load disturbances and process nonlinearities. In order to minimize the effects of these errors, simplified PID tuning formulae are
given in terms of the first three integrated areas (A1 - A3) (Vrani, 1997):
KP =

TI =

0.5
,
A1
K PR
TI

A2 A2 2 4 A1 A3
,
2 A1
TD= TI ,

(2-61)

(2-62)
(2-63)

where is set to = 0.20.25. The advantage of this method is that it allows calculation of the
PID controller parameters merely by measuring the open-loop step response of the process. The
procedure is suitable for on-line calculation of the PID controller parameters.
The objective of the SO method, which was originally proposed by Kessler (1958), is to
obtain an open-loop transfer function of the below formula:

c
2
a c ( s + a )
GO 2 ( s ) = 2
,
s ( s + a c )

(2-64)

where c is the gain crossover frequency and a is related to the phase margin of the control
system through (Peri, 1979, 1998):

2. PID control

27

= 2 atan

FG a 1IJ ,
H a + 1K

(2-65)

or conversely through:
a=

1 + sin
.
cos

(2-66)

The method maximizes the phase margin of the control system and leads to symmetrical
phase and amplitude characteristics, as can be observed in Figure 2.10. The second multiplicand
has the transfer function of a phase-lead network, which provides required phase uplifting at the
frequency c. Figure 2.10 shows the amplitude and gain characteristics of the control system
with the open-loop transfer function equal to (2-64). In the example, the parameter a was set to
4.

|G(j)|
[dB]

40
20
0
-20
-40
10

-2

<G(j)
[deg] -120

c
a

10

-140

-1

ac

10

-160
-180
10

-2

10

-1

10

[rad]

Fig. 2.10. Gain and phase characteristics of a control system tuned according to symmetrical
optimum.
For example, if the process can be modeled with the transfer function G2 (2-3):
G2 (s) =

K2
,
s(1 + T2 s)

it is suitable for the application of the SO tuning method. The procedure leads to a PI controller
with the following settings (Peri, 1979):
KP =

1
,
a K 2 T2

TI=a2T2,

(2-67)
(2-68),

2. PID control
and c =

28

1
. The common choice for the parameter a is 2, which gives the phase margin of
a T2

the control system 370. As with the MO method, the speed of response of the control system
obtained by the SO method is related to the time constant T2 of the process. T2 is considered a
lower time constant, since the integrator in (2-3) can be regarded as the time constant with a very
high value.
The SO method is designed to give a good response to load disturbance, but the response
of the control system to set-point change has large overshoot. The overshoot is commonly
reduced through the usage of a two-degree-of-freedom controller or with a prefilter (strom and
Hgglund, 1995).
The MO and SO tuning methods are widely used in the cascade control systems,
especially to control motor drives (Peri, 1979; Deur, 1999). In such an application, the inner
control loop uses a PI controller, designed according to the MO method, with a measured current
signal as the feedback signal. The outer loop consists of a PI controller tuned using the SO
method and a measured drive speed signal as the feedback signal. As noted previously,
overshoot of the control system is reduced with the prefilter, which compensates the zeros of the
closed-loop transfer function.
In both previous methods (MO and SO), the speed of the control system response is
related to the smaller, non-dominant time constant of the process. Voda and Landau (1995) have
therefore proposed a tuning method based on the knowledge of:
1 the value of the sum of non-dominant time constants (T); and
2 the ratio between the process gain and the dominant time constant.
The method was inspired by the SO method and is termed Kessler-Landau-Voda (KLV) tuning
method.
The region of frequency characteristics of the process where T can be determined is the
region where the phase characteristic equals -1350 (Voda and Landau, 1995). The frequency
where the phase characteristic equals -1350 is denoted with -135, and the process gain at that
frequency is denoted with |G( -135)|. Using these process characteristics, the tuning rules are as
follows:
For the PI controller:
1
,
3.5 G( 135 )

(2-69)

4.6
,
135

(2-70)

4+
,
4 2 2 G( 135 )

(2-71)

KP =

TI =
For the PID controller:
KP =

2. PID control

29
TI =

4+ 1
,
135

(2-72)

TD =

4
1
.
4 + 135

(2-73)

In equations (2-71) (2-73) factor defines the acceleration of response of the control
system using a PID controller in relation to the response of the control system using a PI
controller. It is defined as the ratio between corresponding rise times:

tr PI
.
tr PID

(2-74)

Recommended values for the factor range between 1 and 2.


The KLV method is valid for the type I processes and for the type 0 processes in which
dominant time constant is at least four times as high as the sum of non-dominant time constants.
For type 0 processes that do not satisfy the required ratio between time constants, a control
system tuned according to the KLV method is stable, but has poor dynamic performance. This
tuning method is applicable for on-line usage and autotuning, because it was developed as a part
of an autotuning procedure (Beasanon-Voda and Roux-Buisson, 1997).

2.3.8 PID tuning in the framework of Internal Model Control


Internal Model Control (IMC), thoroughly described by Morari and Zafiriou (1989), is a
general design procedure for obtaining controllers that optimally meet requirements for
stability, performance, and robustness of the control system. Instead of choosing fixed control
structure, and then finding optimal parameters for that structure, IMC postulates a model, states
desirable control objectives, and then finds appropriate controller structure and parameters
(Rivera et al., 1986). The concept of IMC is based on the simulation of the process model GM(s)
within the control structure. Figure 2.11 shows the arrangement of IMC. If the model of the
process GM(s) perfectly matches the process GP(s), and load disturbance is not present, the output
of the model cancels the output of the process annulling thus the feedback signal. In such case,
the process is controlled in an open loop. The feedback signal and hence feedback control, exist
only if there is the model mismatch or load disturbance Z(s). In other words, IMC is a control
strategy in which controller C(s) is designed with the ease of open-loop design, while it retains
the benefits of closed-loop control.

2. PID control

30
GR(s)

Z(s)
+

R(s)
+

C(s)

Y(s)

GP(s)

+
GM(s)

Fig. 2.11. Structure of Internal Model Control.


IMC design is made through following steps. The first is to factor the transfer function
modeling the process:
(2-75)
GM(s)=GM+(s) GM-(s),
+
where GM (s) contains only the left half plane poles and zeros, and GM (s) contains all time
delays and the right half plane zeros. After that, the controller C(s) is defined with:
C(s)=(GM-(s))-1GF(s),
(2-76)
where GF(s) is a filter which guarantees that the controller C(s) is realizable. It also obtains the
desired robustness and defines the closed-loop dynamics. The usual form of the filter is:
GF (s ) =

1
n ,
1 + TF s

(2-77)

but other forms can also be used.


The IMC design procedure can be used to design conventional feedback controllers.
Figure 2.11 shows the relation between a conventional feedback controller GR(s) and IMC
controller C(s), which may be expressed with the below formula:
C( s )
.
1 C( s )G M ( s )

(2-78)

GR ( s )
,
1 + GR ( s )G M ( s )

(2-79)

GR ( s ) =
or inversely:
C( s ) =

For the particular choice of model GM(s) and filter GF(s), IMC design procedure leads to
PID controller (Morari and Zafiriou, 1989). Performance and robustness trade-off of the obtained
control system is handled through the value of the adjustable parameter TF, which determines the
dominant time constant of the closed-loop system. For example, a low value of the parameter TF
produces fast response of the control system, but also results in low robustness margin.
The FODT model (2-2), can be used within the frame of IMC, but the part of the transfer
function modeling dead time e-sTt has to be replaced with Pade approximations (Rivera et al.,
1986). Furthermore, the exponent n in the denominator of the filter transfer function (2-77) is set
to 1. Pade approximation of the zero order is:
e-sTt 1,

(2-80)

2. PID control

31

and leads to an IMC PI controller with the following parameters:


KP =

T1
,
K1TF

(2-81)

TI=T1,

(2-82)

and the recommended value for the filter time constant TF > 1.7 Tt1.
A PI controller that uses a better model of the dead time results in improved behavior of the
control system. It is obtained when the parameters of the controller are set to (Rivera et al.,
1986):
KP =

2T1 + Tt1
,
2 K1TF

TI = T1 +

(2-83)

Tt1
,
2

(2-84)

with the recommended value for the filter time constant TF > 1.7 Tt1.
The first order Pade approximation:
e sTt 1

1 s Tt1 2
,
1 + sTt1 2

(2-85)

in the FODT model and IMC design lead to a PID controller with parameters (Rivera et al.,
1986):
KP =

2T1 + Tt1
,
K1 2TF + Tt1

TI = T1 +
TD =

Tt1
,
2

T1Tt1
,
2T1 + Tt1

(2-86)
(2-87)
(2-88)

and the recommended value for the filter time constant TF > 0.8 Tt1.
PD and PID controllers are obtained as result of IMC design when the process can be
modeled with integration and a time constant (2-3). Parameters of these controllers are:
PD controller
KP =

1
,
K2 TF

TD=T2;

(2-89)
(2-90)

PID controller
KP =

2TF + T2
,
K2 TF 2

(2-91)

TI = 2TF + T2 ,

(2-92)

2T2 TF
.
2TF + T2

(2-93)

TD =

2. PID control

32

Seborg et al. (1989) concluded that the main advantages of IMC design are:
1. Model uncertainty is explicitly considered;
2. Trade-off between performance and robustness of the control system is clearly
defined.
The principal drawback of the method is that the process poles are cancelled with
controller zeros according to (2-76), which results in sluggish response to load disturbance, as
described in subsection 2.3.5.
IMC tuning rules are expressed in terms of process model parameters and can be applied
after the identification of the process model. Such models can be obtained as a part of an
autotuning procedure.

2.4 Usage of PI controller in dead-time compensating controllers


One area of process control in which PID controllers fail to produce satisfactory results is
the control of processes with large time delays. Time delay is considered large when its value
exceeds the dominant time constant of the process. This type of dynamic behavior, termed time
delay or dead time, is present in processes involving transport of materials such as rolling mills
in metal industry and is a common result of composition analysis in chemical industry.
The assertion that PID control is inadequate for the control of processes with large dead
time is based on two arguments:
1. the derivative action of the PID controller, needed for prediction, amplifies noise;
2. the open-loop gain has to be small rendering the performance of the control system poor.
The derivative action of the PID controller can be used to predict, for instance, future
changes of the process output based on the error signal. This type of prediction is called linear
extrapolation and is expressed in the PD part of the PID controller as:
de(t )
),
(2-94)
dt
where e(t) is the error signal. In (2-94), the control signal is calculated using the predicted value
of the error signal e(t+TD) which is presumed to be linear in time. Linear extrapolation can not
be applied to control systems with noise in the measurement signal, since differentiation
amplifies noise. Accordingly, the derivative term of the PID controller is often switched off, and
only the PI part of the controller is employed. However, linear extrapolation can be successfully
applied in temperature control loops, where noise levels are not too high (Hgglund, 1996). The
second problem is that a large time delay introduces a phase lag proportional to the value of the
time delay. As a result, the open-loop gain has to be small in order to preserve stability. Small
open-loop gain system causes slow response and renders the performance of the control system
poor (Laughlin et al. 1987).
u(t ) = K P (e(t + TD )) = K P (e(t ) + TD

2. PID control

33

z(t)
r(t)

Controller

u(t)

GR(s)

y(t)

Process

GP(s)e -sTtp

Model of the process


yM (t+Ttm)
without the delay

Model of the
delay

yM (t)

e-sTtm

GM(s)
+

Fig. 2.12. Structure of the Smith predictor.


Another approach to control of processes with large time delays is to incorporate the
model of the system as in Internal Model Control described in subsection 2.3.8. The Smith
predictor, shown in Figure 2.12, applies that approach. The model of the process is divided in
two parts: one for modeling the dynamic behavior GM(s) of the process and the other for
modeling time delay e sTtm . The transfer function of the control system with the Smith predictor
is:
G X ( s) =

Y ( s)
G R ( s ) G P ( s ) e sTtp
.
=
R( s ) 1 + G R ( s ) G P ( s ) + G R ( s ) G P ( s ) e sTtp G R ( s ) G M ( s ) e sTtm

(2-95)

The denominator of the transfer function (2-95) which is the characteristic equation of the
closed-loop control system, plainly shows that when the modeling is exact (GM(s)=GP(s)) the
two last terms are cancelled. In such case, the closed-loop transfer function becomes:
G X ( s) =

Y ( s)
G R ( s ) G P ( s ) sTtp
=
e ,
R( s ) 1 + G R ( s ) G P ( s )

(2-96)

and the controller GR(s) can be designed as if the process did not contain dead time. In other
words, the controller GR(s) can be designed just for the part of the process modeled by the
transfer function GM(s). The cancellation of dead-time influence on the dynamic behavior of the
control system is characteristic of dead-time compensating controllers.
The main drawback of the Smith predictor is that the performance and the stability of the
control system are very sensitive to inaccurate modeling of the process, especially to the
inaccurate modeling of dead time. Furthermore, it is very difficult to tune the Smith predictor
because it involves precise identification of the process model. Schneider (1988) gave a good
comparison of control performance obtained by pole-cancellation in the PI controller and the
Smith predictor.

2. PID control

34
z(t)

r(t)

KI
s

+
-

u(t)

KM

KR

y(t)

GP(s) e-sTtp

yM (t+Ttm)

1+TMs
+

-sTtm

yM (t) +

F(s)

Fig. 2.13. Structure of the FPPI controller.


One important advantage of PID control over other control strategies, including the Smith
predictor, is that operators are familiar with the tuning procedures for PID controllers. These
involve finding of only three parameters. In comparison, the tuning of the Smith predictor
involves identification of a suitable process model, and then tuning of the controller. The Smith
predictor, which is based on the FODT model and on the PI controller, has five parameters and is
very complicated to tune and operate. In order to simplify the tuning procedure of the Smith
predictor, Hgglund (1996) has proposed a restriction of the choice of the PI controller and
FODT model parameters. This type of the Smith predictor is called the predictive PI (PPI)
controller. Figure 2.13 shows the structure of the PPI controller with the filter F(s) set to 1.
In the PPI controller, the parameters of the PI controller are related to the FODT model
parameters as follows:
KR =

,
KM

(2-97)

1
T ,
(2-98)
M
where and are calculated from the desired performance of the closed-loop control system.
Based on these parameters, the characteristic equation of the PPI is:
TI =

s2 +

1+
s + 2 = 0.
TM
TM

(2-99)

Parameter determines the speed of response (rise time) of the closed-loop system. When is
larger than 1, the closed-loop system is slower, and when is smaller than 1 the closed-loop
system responds more quickly than a system operating in the open loop. When and are set to
1, and when the process exactly matches the FODT model used in the predictor, the closed-loop
system has a double pole at s=-1/TM in the s-plane. Such choice of closed-loop pole positions
ensures aperiodic response of the control system.
Hgglund (1996) compared the performance (IAE) of the classical PI and the PPI
controller in presence of a constant load disturbance. A simple analysis has shown that the
relation between the IAE values corresponding to each controller, is:

2. PID control

35
IAEPPI TM + Ttm
=
.
IAEPI
2Ttm

(2-100)

The equation (2-100) implies that the PPI IAE is 50% lower than the IAE of the classical PI
controller for systems with large time delay, i.e. when Ttm >> TM,. Furthermore, Normey-Rico et
al. (1997) stated that the PPI controller could be interpreted as equivalent to the General
Predictive Controller (GPC), possessing all its advantages.
The PPI controller solves the problem of operational complexity of the Smith predictor
since it decreases the number of controller parameters, but it does not solve the problem of high
sensitivity of the Smith predictor due to the inaccurate modeling of the process. In order to
increase the robustness of the PPI controller Normey-Rico et al. (1997) introduced a filter in the
PPI structure. Figure 2.13 shows such filtered PPI (FPPI) controller. In order to preserve the
simple structure of the PPI filter, F(s) is chosen to be the first-order lag with static gain equal to
one:
1
.
(2-101)
1+ TF s
In Normey-Rico et al. (1997) the robustness of the FPPI controller is analyzed with respect to
different values of the filter time constant TF. The result is the recommendation for its value:
F(s) =

Ttm
.
(2-102)
2
The described approach to control of processes with large time delays is not suitable for
control of processes that have astatic characteristics (processes with type above 0), because
constant load disturbance produces a steady-state error. These processes have to be controlled
with modified Smith predictor structures, suggested in works of Watanabe and Ito (1981), and
strom et al. (1994).
TF =

The described PID controllers are basic components of many control systems operating
in industry. It is worthwhile to automate some of the described tuning procedures. The next step
in the development of autotuning PID controller is to review several variants of the algorithm.
Such review is given in the third chapter.

You might also like