PID CONTROL

Design of PID controllers based on constrained

optimisation
H.Panagopoulos, K.J.Astrom and T.Hagglund
Abstract: A new design method for PID controllers is presented, based on optimisation of load
disturbance rejection with constraints on robustness to model uncertainties. The design also
delivers parameters to deal with measurement noise and set point response. Thus, the formulation
of the design problem captures four essential aspects of industrial control problems, leading to a
constrained optimisation problem which can be solved iteratively.

Introduction

Problem formulation

The PID controller is todays most commonly used control

algorithm [1]. At the moment, there exist many different methods to find suitable controller parameters. The
methods differ in complexity, flexibility, and in the amount
of process knowledge used. Depending on the application,
there is a need to have several types of tuning method.
There are simple, easy to use methods which require little
information, e.g. the method described in [2], as well as
more sophisticated methods which require more information and more computations.
There are several reasons to look for better methods to
design PID controllers. One is the significant impact it may
have because of the widespread use of the controllers.
Another is the benefit emerging auto-tuners and tuning
devices can derive from improved design methods.
This paper describes a new design method for PID
controllers, where the primary design goal is to obtain
good load disturbance responses. This is done by minimising the integrated control error IE. Robustness is guaranteed by requiring that the maximum sensitivity be less than
a specified value Ms . Measurement noise is dealt with by
using filtering. Good set point response is obtained by
using a structure with two degrees of freedom.
The specifications are expressed in terms of a number of
parameters for which good default values can be found. In
the simplest case good default values can be given to all
parameters. The user simply supplies a model of the
process and the design parameter, which is the maximum
sensitivity, Ms . Consequently, the method provides all the
parameters of the PID controller: controller gain k, integral
time Ti , derivative time Td and set point weight b. In
addition, the filters of the measured signal and the set
point are delivered.
For related work, see for example [36].

The design problem is illustrated in Fig. 1. A process with

transfer function G(s) is controlled with a PID controller
with two degrees of freedom. The transfer function Gc(s)
describes the feedback from process output y to control
signal u, and Gff (s) describes the feed forward from set
point ysp to u. Three external signals act on the control
loop, namely set point ysp , load disturbance l and measurement noise n.
The design objective is to determine the controller
parameters in Gc(s) and Gff (s) so that the system behaves
well with respect to changes in the three signals ysp , l and
n, as well as in the process model G(s). Hence, the
specification will express requirements on

# IEE, 2002

where k, ki , kd and b are controller parameters.

It proves beneficial to replace the signals y and ysp with
their filtered values y f and y fsp. The filtered signals are
generated by

IEE Proceedings online no. 20020102

DOI: 10.1049/ip-cta:20020102
Paper received 23rd November 2001
H. Panagopoulos is with the M-real Corporation, Technology Center
rnskoldsvik, O
rnskoldsvik, Sweden
O

K.J. Astrom and T. Hagglund are with the Department of Automatic

Control, Lund Institute of Technology, Box 118, Lund S-221 00, Sweden
32

load disturbance response

robustness with respect to model uncertainties
measurement noise response
set point response

2.1 Process and controller structures

The design problem is formulated so as to apply to a wide
variety of systems. Consequently, the process is assumed to
be linear, time invariant, and specified by a transfer function G(s), which is analytic with finite poles, and possibly
an essential singularity at infinity. The description covers
finite dimensional systems with time delays and infinite
dimensional systems described by linear partial differential
equations.
Initially the controller is described by
t
ut kbysp t  yt ki ysp t  ytdt
0


kd


dyt

dt

Y f s Fy sY s
f
Ysp
s Fsp sYsp s

IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002

l
ysp

Gff

2.4 Measurement noise

In traditional PID controllers the filters are often only
applied on the derivative term. A common choice of
derivative term is

Gc

Fig. 1

Ds 

Conceptual block diagram describing design problem

where the filters are low pass filters of first or second order.
The controller can thus be characterised by four parameters
k, ki , kd , b, and two filters Fy and Fsp .
2.2 Load disturbance attenuation
The primary design goal is to achieve good rejection of
load disturbances. There are no detailed assumptions made
about the load disturbances except that they are lowfrequency. The most common performance criteria is to
minimise IAE, the integrated absolute control error, at load
disturbances. We have used the integrated error IE instead.
The IE criteria is a good approximation of IAE, and the two
criteria are identical for non-oscillatory loops. By combining the IE minimisation with a robustness criterion, welldamped control loops are obtained.
The reason for using IE is that its value is directly related
to the controller parameters, and that it enables simple and
efficient design algorithms. If a unit step load disturbance
is applied at the process input, the IE value becomes [7]:
1
etdt

IE
0

1
ki

The integral gain ki is thus inversely proportional to the

integrated error when a unit step load disturbance is
applied at the process input. By maximising the integral
gain ki the effect of the load disturbance l on the output y is
minimised.

2.3 Robustness
Sensitivity to modelling errors can be expressed as the
largest value of the sensitivity function, i.e.


Ms max
o



1

1 GGc io

The quantity Ms is simply the inverse of the shortest

distance from the Nyquist curve of the loop transfer
function to the critical point 1. Typical values of Ms are
in the range 1 to 2.
The sensitivity can also be expressed as the largest value
of the complementary sensitivity function, i.e.


 GGc io 

Mp max
o 1 GGc io

Typical values of Mp are in the range 1.0 to 1.5.

Another possibility is to use the H1-norm


1 jGGc ioj


g max
o
1 GGc io 
which is discussed in detail in [8].
IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002

kTd s
kd s
Y s
Y s 
Td
k
1s
1s d
N
kN

where N is a number in the range 210. This will reduce

the high-frequency gain to k(1 N). Since the filter is only
applied on the derivative term and not on the proportional
term, the high-frequency gain can never be made smaller
than k.
In this study all terms in the controller are filtered. For
first-order filters with filter time constant Tf , the highfrequency gain becomes kd=Tf . For second-order filters the
high-frequency gain goes to zero as the frequency goes to
infinity. A nice feature of the design is to provide a
systematic way to determine Tf .
2.5 Set point response
The transfer function relating set point to process output is
given by
Gsp s

GGff
F
1 GGc Fy sp
ki bks
GGc
F
2
ki ks kd s 1 GGc Fy sp

When the controller has been designed to give good

attenuation of disturbances, the parameter b and the filter
Fsp can be chosen to give an appropriate set point response.
This is done by determining the maximum of Gsp:
Msp max jGsp ioj
o

2.6 Tuning parameter

The tradeoff between performance and robustness varies
among different control problems. Therefore, it is desirable
to have a design parameter to change the properties of the
closed-loop system.
For the proposed design method the robustness
constraint is a good measure of the performance of the
system. It has been shown in [9] that the variable Ms fulfills
all the requirements of a good design parameter.
3

Difficulties of PID design

The solution of the PID design can be formulated as a

parameter optimisation problem: maximise integral gain ki
subject to the constraints that, (a) the closed loop system is
stable, and (b) the Nyquist curve of the loop transfer
function is outside a circle with centre at s C and
radius R. The constraint (b) can be formulated as
2





C k  i ki  o2 kd Gio
R2
3


o
The constraint that the maximum sensitivity is smaller than
Ms corresponds to C 1 and R 1=Ms . It is also possible
to introduce constraints on the complementary sensitivity
[9]. The solution to the optimisation problem gives good PI
controllers as shown in [9]. For the PID controller the
problem is the lack of constraints, which prohibits the
optimisation of three parameters. This is understood by
33

1.0

ki

0.8
ki

0.6

wR/r

0.4

R/r

0.2
0
0
0.5
k

1.0
1.5

0.5

1.0

1.5

2.0

2.5

3.0

3.5
k

kd

Fig. 2

Geometric illustration of robustness constraint (3). Volume represents parameters for PID controllers such that maximum sensitivity is less
than 1.4. Curve is computed for process G(s) 1=(s 1)4

Fig. 3 Graphical illustration of sensitivity constraint (4)

ranges from 0 to 1 the ellipses have an envelope which

defines one boundary of the sensitivity constraint. Since
the constraint is quadratic in ki the envelope has two
branches; only one branch corresponds to stable closedloop systems. To have a stable closed-loop system it must
also be required that ki is positive.
Fig. 4 shows the envelope for different values of kd . The
Figure illustrates what happens if the integral gain ki is
maximised for a fixed derivative gain kd , where kd 0
corresponds to PI control. It also shows that the integral
gain can be increased substantially by introducing derivative action. Notice that the maximum of ki on the envelope
coincides with the maximum of ki on the locus of the
lowest vertex of the ellipses for small values of kd . This
locus is shown in dashed lines in the Figure. Also notice
that the envelopes have corners for larger values of kd . The
Figure shows that the maximum of ki will typically occur at
a corner of the envelope. When this happens the optimum
is very sensitive to parameter variations. There are also
other drawbacks as will be illustrated by an example.

investigating the robustness constraint in (3). The

constraint defines a set of controller parameters k, ki and
kd such that the maximum sensitivities are less than
prescribed values. An example of such a set is shown in
Fig. 2. The Figure shows that the largest value of ki occurs
at a ridge, which means that the solution is very sensitive to
parameter variations. This phenomenon that occurs for
most transfer functions G(s) illustrates one difficulty of
using derivative action.
3.1 Geometric interpretation
Assuming that the process has positive low-frequency gain,
and introducing
Gio roeijo ao ibo
we find that the constraint in (3) can be written as

2

2
r2
aC
r2
obC
2
k

k

o
k

1
d
r2
r2
R2
o2 R2 i

3.2 Example
Consider the system with the transfer function

For fixed kd and o this represents the exterior of ellipses

with centre in k aC=r2 and ki o2kd 7 obC=r2. The
ellipse has its axes parallel to the coordinate axes. The
horizontal half axis has length R=r and the vertical half axis
has length oR=r. This is illustrated in Fig. 3. When o

1.0

Gs

kd = 0

kd = 1

kd = 2

kd = 3

kd = 3.1

kd = 3.3

1
s 14

0.8
0.6
0.4
0.2
0
1.0
0.8
0.6
0.4
0.2
0
0.5

0.5

1.0

1.5 0.5

0.5

1.0

1.5 0.5

0.5

1.0

1.5

Cuts of robustness region in Fig. 2 for constant derivative gain kd . Curves computed for PID control of process G(s) 1=(s 1)4. Dashed line shows
locus of lowest vertex of ellipses

Fig. 4
34

IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002

Solving the constrained optimisation problem we find that

the controller gains are k 0.925, ki 0.9 and kd 2.86.
Fig. 5 shows the Nyquist diagram of the loop transfer
function. Notice that the Nyquist curve has a cusp. This
will always occur when the maximum occurs at a ridge,
because when there is a maximum at a ridge the Nyquist
curve has a tangency to the sensitivity circle at two
frequencies. This happens when the controller has a
phase lead that increases rapidly, a typical behaviour of a
controller with complex zeros. It is possible to get a very
rapid increase in phase if the relative damping of the zeros
is small. The closed-loop system has two pairs of complex
poles with relative damping z 0.24 and z 0.42, which
means that the time responses can be expected to be
oscillatory.
For comparison we have also designed a controller by
another method which gives the parameters k 1.14,
ki 0.51 and kd 1.14. The Nyquist curve for this controller is shown dashed in Fig. 5. This controller has two pairs
of complex poles with relative damping z 0.59 and
z 0.65, which is more reasonable. The Bode plot for
the loop transfer functions obtained with the controller is
shown in Fig. 6. The Figure shows that the controller that
maximises ki has a large phase lead above the gain crossover frequency.
Fig. 7 shows the responses of the closed-loop system to
steps in the setpoint and the load disturbance. The Figure
shows that the controller which maximises ki gives
responses that are quite oscillatory, indicating that the
design is not too good. The integrated absolute error for
a unit step load disturbance, IAE, gives a quantitative
assessment of the performance. For the controller that
maximises integral gain we find that IAE 3.05 which
can be compared with IAE 2.43 for the controller shown
by a dashed curve in Fig. 7.
The example shows how it is possible to increase
integral gain substantially by the introduction of derivative
action. To obtain large integral gains it is, however,
necessary to introduce a lot of phase lead. By evaluating
time and frequency responses of the obtained closed loop
systems we have found that the large phase lead required to
maximise integral gain has severe drawbacks. In the next
Section we will introduce additional constraints to avoid
these drawbacks.

k = 0.925, ki = 0.9, kd =2.86

k = 1.14, ki = 0.511, kd =1.14

1.0

0.5
_1

0.5

1.0

1.5

2.0
2.0

1.5

1.0

0.5

0.5

Fig. 5

Nyquist curve of loop transfer function for PID control of process

G(s) 1=(s 1)4

k = 0.925, ki = 0.9, kd =2.86

k = 1.14, ki = 0.511, kd =1.14

|L(i)|

101

100

101 1
10

100

arg L(i)

100
120
140
160
180
200
101

100

Fig. 6

Bode plot of loop transfer function for PID control of process

G(s) 1=(s 1)4

step in load disturbance

step in setpoint
0.4

1.0

0.2

1.5

0.5

0.2

2.0

1.5
1.0

1.0

1.5

0.5

0.5
0

10

20

30

40

50

10

20

30

40

50

k = 0.925, ki = 0.9, kd =2.86

k = 1.14, ki = 0.511, kd =1.14

Fig. 7

Time responses for PID control of process G(s) 1=(s 1)4

IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002

35

PID design

Having shown that a direct generalisation of the design

method for PI controllers in [9] is not suitable for PID
controllers, we will now make some modifications. One
possibility is to introduce additional constraints which
prevents excessive phase leads of the controller. After
several attempts, it has been found that good controllers
are obtained by maximising integral gain if the following
constraints are added to the robustness constraint.
v_ w  v w_
<0
_v2 w_ 2 3=2

@ arg Lio
<0
@o
where L(io) is the loop transfer frequency function, with
L(io) v(o) iw(o), where the dot denotes derivatives
with respect to o. The first constraint in (6) specifies that
the Nyquist curve has negative curvature, and the second
constraint prevents the controller from having excessive
phase lead.
The approach will not work for processes that are highly
oscillatory. The second constraint in (6) is not necessary
for processes having integral action or dominating lags.
The systems shown by dashed lines in Section 3 are
obtained for a controller which is designed by using the
constraints in (6).
5

Measurement noise filtering

In controllers with derivative action, the derivative is

filtered by a first-order system. The filter constant is
typically chosen as Tf Td=N where N is in the range of
210. When there is substantial measurement noise, it may
be useful to have filters of higher order, e.g.
Fy s

1
1 sTf n

In the design, it is easy to take the filter into account if it is

applied to all terms of the controller, instead of only the
derivative term which is standard practice. This means that
the filter appears in series with the controller and the
design can be done iteratively.
The choice of the filter time constant Tf in (7) is a tradeoff between reduction of load disturbances and filtering of
measurement noise. A large value of Tf provides effective
filtering of measurement noise, but load disturbances are
not reduced so well. On the contrary, a small value of Tf
gives better reduction of load disturbances, but with less
efficient noise filtering. Simple measures of the effects of

measurement noise are the largest values of the transfer

functions from noise to control signal and to process
variable, i.e.


 Gc io 


Mn max
o 1 GGc io


 GGc io 


Mp max
o 1 GGc io

The maximum complementary sensitivity Mp is a design

parameter and it thus suffices to consider the variable Mn .
For first- and second-order filters it is convenient to
choose the filter coefficient Tf as
8 1
>
>
< No
0
Tf
> 1
>
:
2N o0

for first-order filter

9
for second-order filter

where o0 is the frequency where the sensitivity function

has its maximum. Reasonable values of N are in the
interval 210. In Table 1 the influence of the filter on the
loop transfer function at the critical frequency o0 is shown
by calculating the arg Fy(io0). Note that the special choices
of Tf in (9) for the first- and second-order filters will give
approximately the same amount of influence of the loop
transfer function at o0 for a certain value of N.
The insertion of a filter will modify the loop transfer
function, but with suitable values of N these changes are
kept small. Nevertheless, it is possible to take these
changes into account by repeating the design with process
G replaced by FyG. The results of such a repeated design
are shown in Table 1, where the process is given in (5).
The trade-off between filtering capacity and loss of
control performance is demonstrated in the Table. A
large value of Mn is obtained when the IAE value is
small. The advantage of the proposed design method is
its systematic way to determine measurement noise filters,
such that the noise level is reduced.
6

Set point response

To obtain a complete design it remains to find a suitable

value of the set point weight b. One possibility is to
consider the transfer function from set point to output,
and to make sure that the maximum magnitude of this
transfer function is not larger than 1. This gives a set point
response without overshoot for most systems. For a

Table 1: Properties of PID controllers obtained for system FyG with G(s) 1/(s 1 1)4 and Ms 1.4 for different filter of
order n 1, 2
n

10

10

arg Fy(io0)

50.

12.57

6.35

59.

12.69

6.36

0.

0.7895

0.9152

1.0009

0.6496

0.8161

0.9671

1.140

Ti

2.8334

2.3322

2.2867

2.8285

2.4033

2.2963

2.227

Td

1.2979

1.0406

1.0130

1.3083

1.0716

1.0180

0.999

o0

0.5091

0.6695

0.7227

0.4703

0.6226

0.7011

0.790

IAE

4.2425

3.0711

2.7821

4.9524

3.4774

2.8825

Mn

1.1392

3.7584

7.9840

1.0726

2.0550

6.1308

36

2.4209
1

IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002

controller with setpoint weighting, the response to setpoint

changes is given by the transfer function
Gsp s

GsGff s
1 GsGc sFy s

10

ki bks
GsGc s

ki ks kd s2 1 GsGc sFy s
A straightforward search procedure can be used to determine if there is a value of b that makes the maximum of
jGsp(io)j equal to 1. Such a procedure will give a suitable
value of b if one exists.
Only nonnegative values of b are allowed, since negative
values may result in inverse step responses in the control
signal, which is an undesirable way to reduce overshoots. If
jGsp(io)j is larger than 1 for b 0, a low-pass filtering of
the set point may be used to reduce the magnitude of
jGsp(io)j further.
The set point filter Fsp(s) can be determined in the
following way. Let ms be the maximum of the transfer
function in (10), and let osp be the frequency where the
maximum occurs. A first-order filter
Fsp

1
1 sTsp

has the magnitude 1=ms at the frequency osp if the time

constant is
1 p
Tsp
m2s  1
osp
Feeding the set point through a low-pass filter designed
in this way will reduce the magnitude at the frequency
osp to 1.
7

Numerical solution of design problem

Finally, a brief discussion of the numerical solution of the

optimisation problem for the design of PID controllers in
(6) is given. The optimisation problem is non-convex and
the solution is obtained numerically with the Optimisation
Toolbox in Matlab 5, which requires substantial calculations. However, with todays personal computers there are
no major limitations.
As for most numerical optimisation routines it is important to have good initial conditions and a suitable search
interval. A natural choice of initial conditions will be the
controller parameters k and ki from the PI design in [9],
that is,
k0

ki0

kd0 k

k i

and a suitable search interval is given by

 0 =2
ostart o
 180 o
 270 =2
ostop o
o
0 is the frequency at which the sensitivity function from
the PI design has its maximum. o
270 are the
180 and o
frequencies where the argument of the loop transfer function from the PI design is 180 and 270 , respectively.
The following procedure is used to solve the design
problem:
(i) Obtain a model of the process in terms of a transfer
function.
(ii) Choose the design parameter expressed by the maximum of the sensitivity function Ms .
IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002

(iii) Determine the number of constraints, as it differs

depending on the considered system.
(iv) Make a PI design to obtain the initial values [k k i 0]
and the frequencies [o
180 o
270].
0 o
(v) Solve the design problem with the Optimisation Toolbox in Matlab 5.
(vi) Verify that the resulting controller parameters fulfil
the constraints. If not, adjust the initial values or settings in
the accuracy of the numerical routine.

Examples

The design method has been tested on a number of

examples which illustrate its properties. The following
transfer functions have been considered:
e5s
s 13

G1 s

1
;
ss 13

G3 s

1
s 11 0:2s1 0:04s1 0:008s

Gn s

1
; n 4 to 7;
s 1n

G2 s

G8 s

1  2s
s 13

The first seven models capture typical dynamics encountered in the process industry. Model G1 is an integrating
process and G2 models a process with long dead time.
Model G8 has a zero in the right half-plane, which is
uncommon in process control, but it has been included to
demonstrate the wide applicability of the design procedure.
Figs. 8 and 9 show the responses to changes in set point
and load. The details of the design calculations and
simulations are summarised in Table 2. Note that the PID
controller obtained is compared to the corresponding PI
controller to show the amelioration of the PID design.
Although models G1G8 represent processes with large
variations in process dynamics, Figs. 8 and 9 show that the
resulting closed-loop responses for a load disturbance
become similar for each value of Ms . This is important
because it means that the proposed design procedure gives
closed-loop systems with desired and predictable properties.
There is also a clear similarity between the responses
obtained with the different values of the tuning parameter
Ms , which indicates its suitability as a tuning parameter.
Responses obtained with Ms 1.4 show little or no overshoot, as is normally desirable in process control. Faster
responses are obtained with Ms 2.0. The settling time at
load disturbances, ts , is significantly shorter with the larger
value of Ms . On the other hand, these responses are
oscillatory with larger overshoots. This can be seen from
the comparison between IE and the integrated absolute
error IAE in Table 2.
The controller gain k varies significantly with the design
parameter Ms: it is larger for designs when Ms 2.0 than
for those when Ms 1.4. However, integral time Ti is fairly
constant for the stable processes, i.e. all processes except
G1 . The derivative time Td is usually larger for designs
with Ms 1.4 than for those with Ms 2.0. In all cases the
PID design generates a controller with complex zeros for
Ms 2.0. Thus, the controller will not be realisable in
serial form.
For Ms 2.0 the Mp values are large. Consequently, the
overshoots would be significant if the set point weight is
b 1. However, acceptable set point responses are
obtained by suitable choices of either the set point
37

1.5

2.0
1.0

1.0

1.5

0.5

0.5

0.10
0.05
0
0.05
0.10

1.5

0.5
0
50

100

150

200

250

20

40

60

80

100 120 140 160

b
1.5

1.0

1.0

1.5

y
0.5

0.5

10
8
6
4
2
0
2

1.5

1.0

1.0
0.5
0

10

30

20

40

50

60

1.5

1.5

1.0

1.0
0.5

1.5

1.5

1.0

1.0

0.5

0.5

0.5

0
0

10

20

30

40
e

50

60

70

80

1.5
1.0

1.5

0.5

0.5

1.5

1.5

1.0

20

30

50

40

60

70

80

90

1.0

0.5

10

1.0
0.5

0
0

10 20 30 40 50 60 70 80 90 100 110
g

10 20

30 40 50
h

60 70 80 90 100

Fig. 8 Comparison between PID (solid line) and PI controller (dashed line) for Ms 1.4, showing step response followed by load disturbance of closed
loop system
a system G1
b system G2
c system G3
d system G4

38

e system G5
f system G6
g system G7
h system G8

IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002

1.0

1.0

1.5

0.5

0.5

0.6

1.5

0.4

1.0

y
u

1.5

0.2

0.5

0
0.2

0
50

100

150

40

20

60

a
1.5

1.0

120

1.0
0.5

30

1.5

20
10

100

1.5

0.5

1.0
0.5

0
10

80

0
0

10

20

40

30

50

60

70

d
1.5

1.0

1.0

1.5

0.5

0.5

1.5

1.5

1.0
0.5

1.0
0.5

0
0

10

20

30

40

50

60

70

1.5

20

30

50

40

60

70

80

90

2.0
1.5
1.0
0.5
0
0.5

1.5

1.5

1.0

1.0

1.0

10

0.5

0.5

0.5

0
0

10 20 30 40

50 60 70 80 90 100
g

10

20

30

40

50

60

70

Fig. 9 Comparison between PID (solid line) and PI controller (dashed line) for Ms 2.0, showing step response followed by load disturbance of closed
loop system
a system G1
b system G2
c system G3
d system G4

e system G5
f system G6
g system G7
h system G8

IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002

39

Table 2: Properties of obtained controllers for system G1G8 with different values of design parameter Ms
Process

Ms

G1(s)

1.4
2.0

G2(s)
G3(s)
G4(s)
G5(s)
G6(s)
G7(s)
G8(s)

Td

Tsp

IE

0.324

6.59

2.35

0.00

0.17

20.4

0.680

4.50

2.27

0.00

0.08

1.4

0.325

3.55

1.69

0.69

0.00

2.0

0.555

3.21

1.74

0.00

1.61

ts

Mp

0.67

0.77

15.3

1.53

0.69

0.91

37.8

1.50

0.99

0.26

48.2

1.00

5.80

0.64

0.29

39.2

1.32

6.61
10.9

15.96

0.212

0.15

0.00

0.26

0.013

0.57

19.1

2.99

1.52

2.0

43.13

0.189

0.13

0.81

0.00

0.0044

0.75

25.6

3.35

1.65

1.4

1.14

2.23

1.00

0.00

0.27

1.95

0.81

0.79

17.4

1.09

2.0

2.27

1.91

0.98

0.00

0.53

0.84

0.61

0.95

15.3

1.62

1.4

0.784

2.68

1.24

0.00

0.51

3.41

0.84

0.54

23.9

1.04

2.0

1.47

2.33

1.25

0.00

0.72

1.59

0.56

0.63

21.0

1.55

1.4

0.635

3.12

1.47

0.00

0.44

4.92

0.88

0.42

27.5

1.01

2.0

1.15

2.74

1.49

0.00

0.97

2.38

0.56

0.48

23.3

1.50

1.4

0.552

3.57

1.69

0.00

0.25

6.44

0.90

0.34

32.7

1.00

2.0

0.982

3.14

1.73

0.00

1.24

3.20

0.57

0.39

28.2

1.47

1.4

0.312

2.25

0.80

0.60

0.00

7.22

0.86

0.51

30.5

1.00

2.0

0.542

2.07

0.79

0.00

1.03

3.82

0.62

0.58

22.4

1.31

Conclusions

PID controllers were designed to capture demands on load

disturbance rejection, set point response, measurement
noise and model uncertainty. Good load disturbance
responses were obtained, minimising the integrated control
error IE. Robustness is guaranteed by requiring a maximum sensitivity of less than a specified value Ms . Measurement noise is dealt with by filtering. Good set point
response is obtained by using a structure with two degrees
of freedom.
The design procedure has been applied to a variety of
systems, stable and integrating, with long dead times and
with right half-plane zeros. The method can also be used to
develop simple tuning rules which are significant improvements of the ZieglerNichols rules [10].

40

o0

1.4

weight b or the filter Fsp . According to Table 2, it is not

always enough to set b 0 to obtain a small overshoot;
filtering may also be needed.

IE=IAE

Ti

10

References

YAMAMOTO, S., and HASHIMOTO, I.: Present status and future

needs: the view from Japanese industry. Proceedings of fourth international conference on Chemical process control, Texas, USA, 1991
2 ZIEGLER, J.G., and NICHOLS, N.B.: Optimum settings for automatic
controllers, Trans. ASME, 1942, 64, pp. 759768
3 SCHEI, T.S.: Automatic tuning of PID controllers based on transfer
function estimation, Automatica, 1994, 30, (12), pp. 19831989
4 VAN OVERSCHEE, P., and MOOR, B.D.: Optimal PID control of
chemical batch reactor. Proceedings of 1999 European control conference, Karlsruhe, Germany, 1999
5 LANGER, J., and LANDAU, I.D.: Combined pole placement=
sensitivity function shaping method using convex optimization criteria,
Automatica, 1999, 35, (6), pp. 11111120
6 KRISTIANSSON, B., and LENNARTSSON, B.: Optimal PID controllers including roll off and Schmidt predictor structure. Proceedings of
14th world congress of IFAC, Beijing, P.R. China, 1999, Vol. F, pp. 297
302
STRO
M, K.J., and HA
GGLUND, T.: PID controllers: theory, design
7 A
and tuning (Instrument Society of America, North Carolina, 1995)
STRO
M, K.J.: PID control design
8 PANAGOPOULOS, H., and A
and H1 loop shaping, Int. J. Robust Nonlinear Control, 2000, 10,
pp. 12491261
STRO
M, K.J., PANAGOPOULOS, H., and HA
GGLUND, T.: Design
9 A
of PI controllers based on non-convex optimization, Automatica, 1998,
34, (5), pp. 585601
STRO
M, K.J., and HA
GGLUND, T.: New tuning methods for PID
10 A
controllers. Proceedings of European control conference, Rome, Italy,
1995, pp. 24562462

IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002