PI0 CONTROL

PID controllers: recent tuning methods and design

to specification

P.Cominos and N.Munro

Abstract: PID control is a control strategy that has been successfully used over many years.
Simplicity, robustness, a wide range of applicability and near-optimal performance are some of the
reasons that have made PID control so popular in the academic and industry sectors. Recently, it
has been noticed that PID controllers are often poorly tuned and some efforts have been made to
systematically resolve this matter. In the paper a brief summary of PID theory is given, then some
of the most used PID tuning methods are discussed and some of the more recent promising
techniques explored.

1 Introduction form to the parallel form and vice versa under certain
conditions. It is also very important to know the controller
A PID controller, Fig. 1, is described by the following implementation before entering the PID parameters.
transfer function in the continuous s-domain In cases where both good set-point tracking and good
load disturbance are essential, the set-point weighting
G ~ ( s=) P + I + D = Kp +'+K . S
K ~ s (1) technique [l] can be used. This method effectively results
in a two-degrees-of-freedom (2DF) system. Every input to
or each of the PID terms can be a weighted version of the
error signal. Set-point weighting is not used with propor-
tional-only control since it results in static control errors.
The general 2DF PID configuration is shown in Fig. 2.
where Kp is the proportional gain, Ki is the integration The ISA realisation is a particular case of a 2DF
coefficient and Kd is the derivative coefficient. T, is known configuration. Here, the controller takes the form
as the integral action time or reset time and Td is referred to
as the derivative action time or rate time. G&) = GSP(S>YSP(S)- Gs(S)Y(S) (6)
In practice, the following realisation is usually employed
where GsP(s) is the signal transmission from the set-
Ki K ~ s point to the control variable and Gs(s) is the signal
GJs) = Kp + s + 1 + T,s
- ~
(3) transmission from the process output to the control
variable [2].
This realisation, although it has the same response at low
frequencies as (l), includes a lowpass filter with the
derivative term to reduce noise amplification. Tn represents
the filter's time constant.
The PID realisation given by (3) is known as the parallel
form and admits complex zeros. One can also have the
series PID form described by
GcG) = GPl(S)GPD(S) (4)
or
I I
Fig. 1 Block diagram representation of a PID controller

It is also required that p < 1 to obtain the phase lead action.

This configuration is less flexible than the parallel form
due to a stronger interaction between the design para-
meters. Note, that it is possible to convert from the series

0 IEE, 2002
IEE Proceedings online no. 20020103
DOI: 10.1049/ip-cta:20020103
Paper received 7th January 2002
The authors are with the Control Systems Centre, UMIST Sackville Street,
Manchester M60 lQD, UK Fig. 2 2DF PID controller
46 IEE Proc.-Control Theory Appl.. Vol. 149, No. I, January 2002
 The remainder of this paper describes different €'ID
parameter tuning methods together with a discussion on
some of their advantages and disadvantages.
2 Ziegler-Nichols (ZNI method
The PID tuning method presented by Ziegler and Nichols
[3] is based on the system's open-loop step response. It
uses the fact that many systems in the process industry can
be approximated by a first-order lag plus a time delay as
where c( and L can be determined by simply plotting the
step response of the plant [l]. The PID tuning parameters
obtained by the ZN step response method are shown in
Table 1.
Ziegler and Nichols later introduced a method based on
.the frequency response of the closed-loop system under
pure proportional control. Here, the gain is increased until
the closed-loop system becomes critically stable. At this
point the ultimate gain, K,, is recorded together with the
corresponding period of oscillation, Tu, known as the
ultimate period. Based on these values Ziegler and Nichols
calculated the tuning parameters shown in Table 2.
The ZN methods were designed to give good responses
to load disturbances. A quarter amplitude-damping criter-
ion was used in the design giving a damping ratio close to
0.2. This is not satisfactory for many systems, since it does
not give satisfactory phase and gain margins. The maxi-
mum sensitivity is also large, giving systems sensitive to
parameter variations. Additionally, ZN methods are not
easy to apply in their original form on working plants.
When critical processes are involved, sudden changes in
the control signal or operation at the stability limit are not
acceptable. Relay feedback and describing function analy-
sis [l] are often applied for parameter identification to
overcome the above problems.
A further modification to the ZN methods can give a
substantially improved system performance [ 11. One can
start with a given point on the Nyquist diagram say,
G,( jo)= rpe'(K+9(3 and then try to find a regulator to
move this point to B=r,,e'("'4'). An amplitude margin
corresponds to 4% = 0, r = 1 /Am and a gain margin corre-
sponds to r, = 1, 4, = I$,, , where A , , (bm are the gain and
phase margins, respectively. Based on this simple modifi-
cation, a better system response results. If the plant's model
is well known then it is also possible to apply Routh's array
technique to find the critical gain and then the correspond-
Table 1: ZN PID step response tuning parameters
Controller K T, Td
P 1l a
PI 0.9la 3L
PID 1.21E 2L u2
Table 2: ZN PID frequency response tuning parameters
Controller K Ti Td
P 0.5Ku
PI 0.4K, 0.8Tu
PID 0.6K, 0.5T,, 0.12T,
IEE Proc.-Control Theoiy Appl., Vol. 149, No. 1. January 2002
ing critical period. In [ 11, area methods are also suggested
to improve estimation of the system's parameters giving a
less sensitive technique.
3 Kappa-tau tuning
The dynamics of a system can be described more accu-
rately if three parameters are used in the design instead of
two. The kappa-tau tuning method [ 1,4] is a method in that
direction and is used in automatic tuning. As in the ZN
method it comes in two versions. One is based on the step
response, in which the process is characterised by a static
gain Kp, a gain a (the gain of the transient part of the open-
loop response), and a dead time L. The controller para-
meters are a function of the normalised dead time z [4, 51
given by
L
z=-
L+T
with T being the dominant time constant of the process.
The second method is based on the frequency response;
the process is characterised by a static gain K p , an ultimate
gain Ku and an ultimate period Tu. Here, the controller
parameters are a function of the gain ratio k, [4, 51, where
Maximum sensitivity is used as the design objective in
both cases.
4 Pole placement
Analytical pole placement methods are mostly used when
the system under consideration is of low order. A common
approach is to adopt a second-order model and then specify
a desired damping ratio and natural frequency for the
system. These specifications can then be fulfilled by
locating the two system poles at positions that give the
required closed loop performance. As an example, the
characteristic equation for a system approximated by a
first-order model
G~(s) KP
~
(10)
1+ST,
under PI control will take the form:
This can then be compared with the general second-order
model
s2 + 2go + o2= 0 (12)
and thus obtain
In the case where a second-order system of the following
form is used:
a PID controller of the form
47
can arbitrarily place all closed-loop poles. The system’s
characteristic equation then becomes
This can be compared with the following, general, third-
order characteristic equation [4]:
(s + Cto)(s2+ 2q3s + 0 2 ) = 0 (17)
to get the PID parameters, as for the PI case earlier.
Dominant pole design is another, simplified, pole place-
ment technique employed when one wants to obtain a PID
controller for high-order systems. This method is based on
the positioning o f the system’s dominant poles in the
complex plane. For example, by taking the transfer func-
tion of a unity feedback system
one can then readily find the poles and zeros of the
resulting closed-loop system. In many cases, the dominant
system dynamics can be approximated by the simple pole-
zero configuration shown in Fig. 3. The pair of poles P I ,
P2 are known as the dominant poles. Poles and zeros which
have real parts much more negative than those of the
dominant poles have little influence on the overall system
response. For a PI controller, as an example, the following
selection of P I , P2 can be made [l]:
These give a system with the required damping c, and
response speed given by coo.
For an extensive analysis of pole placement techniques
with emphasis on uncertain systems, the reader is referred
to [6].
5 Design based on gain and phase margin
specifications
Usually one wants to design for specific gain and phase
margins [7], where the phase margin is related to the
damping of the system. By definition, the gain margin of
t’.
Fig. 3 Dominunt poles
48
a system is given by the solution of the following set of
equations:
= --7c
arg[GdJwP)G~(J~,>l (20)
Correspondingly, the phase margin is given by
where cop is the phase crossover frequency, cog is the gain
crossover frequency, and Gp(s), Gc(s) are the process and
the controller transfer function, respectively. It is now
apparent that, depending on the plant’s model, the solution
of the above sets of equations can be extremely difficult to
carry out analytically. Numerical methods are usually
employed.
In [8], a first-order system having the form
Gp(S)
KP e-sL
=-
1 +sT
is controlled with a PI controller having the following
transfer function
By trying, in this case, to solve for a specific phase and
gain margin one comes across equations containing arctan
functions, which cannot be solved analytically. This
problem is overcome by using the following simple
arctan approximations
6 D-partitioning
In the D-partitioning method [9], the system’s poles are
shifted into a specified region in the left halfplane. Least
absolute stability is initially achieved by placing all the
system’s poles to the left of a line parallel to, and at a
distance d from, the imaginary axis. Then, a region with
this line as its right boundary is obtained within which the
specific gain and phase margins are guaranteed. Therefore,
relative stability criteria are also met.
In [lo], a detailed description of how to use the
D-partition method for PID tuning is given. In addition,
a method of shifting the closed-loop system zeros to
specified positions in the s-plane is proposed to improve
the system’s tracking behaviour. Regions of absolute and
relative stability in terms of the PID tuning parameters are
obtained.
7 OLDPmethod
The OLDP method [ 1 11 is an extension of the D-partition-
ing method, where simultaneous achievement of minimum
phase and gain requirements is possible. In addition, this
method allows for specified maximum gain and phase
crossover frequencies. This is possible by isolating the
IEE Proc.-Control Theory Appl., Vol. 149, No. I , Januuy 2002
gain and phase margin regions in parameter space. A plant This theorem can be further extended to what is known as
model with real coefficients of the following form is the generalised Kharitonov theorem [12, 131, which plays
assumed: an important role in the area of parametric control design.
The generalised Kharitonov theorem states that:
1. Given an m-tuple of fixed real or complex polynomials
of the form [ F ,(s), F2(s),. . . ,Fm(s)],the polynomial family
with c, # 0. The open-loop transfer function for the PID described by
case is then given as
PI @)FI
(s) + P2(s)F2(s)+ . . . + P,(s)F,(s) (34)
+
Kds2 Kps + Ki
G(s) =
S
Gp(s) = a +jb (28) with Pi(s) an interval polynomial, is Hunvitz stable if and
only if all of the one-parameter polynomial families that
s is then substituted by j w and the characteristic equation result from replacing [Pl(s),P2(s),. . . , P,(s)], with m an
becomes arbitrary integer, in (34) by elements S, (the family of m4m
distinct generalised Kharitonov segments), are Hunvitz
-Kdw2 +j K p o + K, - R(w) - jI(cu) = 0 (29) stable.
2. If the polynomials Fi(s)are real and of the form
where R(o), I ( o ) are the real and imaginary parts of +
Fi(s)=stfais bj)Uj(s)Qi(s),where ti 2 0 is an arbitrary
jw(a+jb)/Gp(s). To get the required gain and phase integer, a i and bi are arbitrary real numbers, Ui(s) is an
margins in parameter space, three frequency domain anti-Hunvitz polynomial, and Qi(s) is an even or odd
boundaries have to be drawn. These boundaries are polynomial, then it is sufficient that (34) be Hunvitz
obtained by considering the following three cases: stable with all Pi replaced by the elements of K,. K, is
1. o = o the set of m-tuples obtained as follows: for every fixed
2. o = f o o , and integer i between 1 and m set Pi@)=Kf(s)for some k = 1,
3. O < o < c o a n d O > w > - o o . 2, 3, 4.
3. If the Fi are complex and
For the PI case, the characteristic equation takes the
following form: d
-arg[Fi(jw)] 5 0, V i = 1 , 2 , . . . , m
jKpo + K, - R(w) - jZ(o) =0 (30)
do
then it is sufficient that (34) be Hunvitz stable with the P,
When w = 0 the corresponding boundary is given by replaced by the elements of K, .
cmK,=O, which implies that K,=0. In the case where
o= f c o , the boundary is described by The generalised Kharitonov theorem can been seen as an
+
Kpco(jw)'n+' - (a jb)(jco),,+' = 0, and finally when improvement on the edge theorem [ 141. The edge theorem
0 < o < oo and 0 > o > -co the boundaries are given by considers polynomials with affine linear uncertainty struc-
K, =R(o)and Kpw =I(@). ture; i.e. their coefficients are dependent on each other. The
The PID case can be similarly analysed but here one of edge theorem states that a family of polynomials is stable if
the K,, Kd parameters has to be gridded. the exposed edges of the polytopic family are stable. In the
Points on the boundaries specified above are related to generalised Kharitonov theorem, the number of edges
the gain and phase crossover frequencies and thus the which have to be studied for stability depends on the
bandwidth of the systems under consideration. In particu- number of interval polynomials and not on the number
lar, a point on the non-singular phase margin boundary of uncertain parameters, making it a computationally more
corresponds to a specific frequency, the gain crossover efficient method.
frequency of the controlled system. Similarly, a point on In [ 151, the above ideas are combined with results in the
the non-singular gain margin boundary corresponds to the area of robust parametric control and the generalised
phase crossover frequency. Hermite-Biehler theorem [ 161 to obtain all stabilising P,
PI and PID controllers for an interval plant family. Bulut
[17] has extended some of the above results. In addition to
8 Interval polynomial techniques stability, he obtained PID controllers for uncertain systems
that also met system performance requirements.
An interval polynomial is a polynomial of the following In [ 181, bilinear transformation and linear programming
form: together with the above results are used to determine the
p(s) =po +p1s+p2s2 + . . . + p ns" set of all PID gains that can stabilise a given discrete-time
(3 1)
plant of arbitrary order.
with the coefficients varying in independent intervals given
by 9 Nyquist based design
PO E ~ n[ ~ n ~ ~ n l ~ 0 ~ [(32)
[ x o > . ~ o I > . . . >E ~ n ~ ~ n l
In [19], Munro introduced some fast ways of system-
Then, Kharitonov's theorem states that every polynomial in atically calculating the limiting values of PID parameters
the interval family is Hunvitz stable if and only if the based on the axis crossing form of the Nyquist's stability
following four polynomials are Hunvitz stable
theorem. Namely,
K 1 ( ~ ) = x o + x , s + y 2 s 2 + y 3 s 3+x4s4+ Z=N+P (35)
K2(s) = xo + y l s +y2s2 + x3s3 + x4s4 + where Z is the number of closed-loop poles located in the
(33) right halfplane, P is the number of the open-loop poles of
K3(s)=y~+xls+x*s2+y3S3+,v4S4+"' the system located in the right halfplane and N is the
number of poles crossing the imaginary axis at the point
K4(s) =yo + y l s +x2s2 +x,s3 +JJ4S4 + ... where the Nyquist plot passes through the -1 point. For
IEE Proc.-Control Theory Appl, Vol 149, No 1, Junuury 2002 49

test
K
comp. plant
Fig. 4 Test compensator arrangement
system stability, one wants Z= 0. The system configuration
shown in Fig. 4 is used.
Values of K that make each real axis crossing equal to
the critical point on the Nyquist diagram are obtained and
then used to scale the test compensators. This procedure is
done by gridding one of the PID parameters and recording
the limiting values of the other two. For systems with no
explicit time delay, the procedure can be simplified further
by simply solving Im[G(s)K(s)] =0, since only the real
axis crossings on the Nyquist diagram are required. The
method is then further improved giving a computationally
efficient way of obtaining the limiting stabilising values of
the PID controller. It can be proved [20] that when the Kp
tuning parameter is gridded then the others are confined
within convex polygons. Linear programming techniques
can then be used to obtain the vertices of these polygons.
This method can be applied to stable systems, unstable
systems, minimum phase and non-minimum phase systems
and systems with time delay. Ideas from [15, 211 are then
employed to handle systems with structured parameter
uncertainty resulting in robust PID controllers.
10 Genetic algorithms for PID tuning
Genetic algorithms are a rapidly expanding area in control
systems design. A genetic tuning algorithm usually starts
with no knowledge of the correct solution and depends on
the responses from its environment to give an acceptable
result. It has been shown that genetic algorithms are
capable of locating optimal regions in complex domains
avoiding the difficulties, or even erroneous results in some
cases, associated with the gradient descent methods and
with high-order systems. To obtain the PID tuning para-
meters one usually has to minimise a performance index.
This, in the majority of the cases, is one of the following:
ISE =
I" I r(t) - y(t)I2dt
IAE =
fi I ~ ( t-)~ ( tI dt)
ITAE =
$I tl r(t) - ~ ( tI dt)
with r(t) being the reference input and y(t) the output of the
system.
In [22], a genetic algorithm based on Gray coding is
used. Each PID parameter ( K p ,K,, Kd) is represented by 16
bits and a single individual is generated by concatenating
the coded parameter strings. The genetic algorithm requires
a population of initial approximations, which may be
random, to start the search. The algorithm then checks
the fitness of each individual (or chromosome), and then
grades them. A selection process follows where five of the
fittest individuals are chosen. The remaining individuals
are selected probabilistically. The selected individuals are
50
used to produce the next population, and the process is
then repeated until the design requirements are met. This
method is applicable to a wide range of system models due
to its adaptability. High-order systems do not present a
problem with this tuning procedure.
11 PID tuning using the theory of adaptive
interaction
In [23], an adaptive method based on the theory of adaptive
interaction [24] is used for tuning PID controllers. Tuning
is obtained through minimisation of a performance index
(e.g. the error squared). The controlled system is broken
down into four subsystems; namely, the plant, the propor-
tional control, the integral control and the derivative
control, K p , K, and K d are viewed as the interaction
parameters between the four subsystems. An adaptive
interaction algorithm is then used to tune these parameters.
For this purpose, the system's Frechet derivative [25] is
needed. In many cases, the Frechet derivative can be
replaced by a constant that will be absorbed in an adapta-
tion coefficient. This gives flexibility in the design, since it
reduces the dependency on the plant's model. The K p , K,,
Kd parameters are treated as variables and they are not
required to converge to a constant value. In fact, they
change as the set-point and the disturbances entering the
system change.
The theory of adaptive interaction is based on the
assumption that a complex system is composed of other
subsystems called devices. The dynamics of each device
are given by a functional (a transformation from a vector
space X into the space of real (or complex) scalars is said
to be a functional on X) F,: xn +y n , n E K, where x,, yn
denote the input and output spaces, respectively. In other
words, the output of each device is a function of each
input, which in turn is related to the outputs of other
devices. Additionally, it is assumed that the input to a
device is a linear combination of the output of the other
devices via connections in I,, the set of input interactions
for the nth device, and possibly an external signal u,(t)
described by x,(t) = u,(t) + ~ c c , ~, ~ y ~ , .n~E ~K,( where
t),
a, is a connection weight, and ypre, is the output of another
device whose output is conveyed by connection c. The
dynamics of the system can now be described by y,(t) =
+
Fn[un(t) C, a,ypre,(t)], n E K. The target is to find the
correct a, so as to minimise some performance index
E(.YI,.. . ,yn,u i , . . .u,>.
For the PID case, the system is decomposed into four
subsystems as explained earlier. Then, the adaptive algo-
rithm requires that
(36)
where the o symbol denotes composition, i.e.
In [23], the error squared is minimised as the performance
index giving the following auto-tuning parameters:
Kp = -2yeF:[x4] o y l
Ki= -2ye~:[x,] o y2 (39)
Kd = -2yeFi[x4] o y 3
IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002
Then, the Frkchet derivative is approximated by present [l], it is not recommended for some cases [ 3 2 ] .
F‘[x] 0 h = ph, with p being a constant. This can then be Good set-point responses may be achieved, but load
substituted into (39) to get disturbance response can be poor. In the case where
unstable poles are involved, then cancellations must be
K,] = -YeYl avoided because they will give an unstable response.
In [26], a simple but effective order reduction technique
Ki = -yey, (40) for use under PID control action is introduced. The plant’s
model is initially reduced by retaining the slowest poles,
Kd = -WY, and then by retaining the low-order coefficients. It is then
The advantage of such a method is that it can be applied to proven that the former is an overestimate of the plant’s
both open-loop stable and unstable plants, including magnitude, whereas the later is an underestimate. By
systems with time delay, and it is robust to changes in taking the average of these two, a reasonably good approx-
the system’s parameters and/or to noise entering the imation of the plant’s model is achieved up to its bandwidth
system. point. In [33], the results of [26] are extended to deal
with discrete-time systems. The main advantage of such
methods is their simplicity. For an extensive reference on
12 Methods based on cancellation model order reduction techniques, the reader is referred to
[34, 351.
It has been pointed out in [26, 271, that PID control is often
based on a second-order model. This implies that when
PID action is used for control, then a model order reduc- 13 K-B parametrisation
tion has to be employed. Internal model control (IMC) [28,
291 is useful in this case, since a plant model order K-B, Keviczky-Banyaz, parametrisation is a generic two-
reduction automatically results in controller order reduc- degree-of-freedom (G2DF) implementation. It provides the
tion. All stabilising controllers of a system under IMC are desired transient responses considering both tracking and
parameterised as regulatory performances, and it effectively opens the
system’s closed loop. Figs. 5 and 6 illustrate the principle,
where S EI, Ho E 1 and Ro are the process model, the noise
model and the controller, and 1 is the set of all stable proper
systems.
where ,G(s) is the system’s transfer function and
In [36], classical PID tuning techniques are combined
G(s)= G-(s)G+(s). Here, G+(s) includes all non-minimum
with the K-B parametrisation method to give improved
phase dynamics and Q(s) is usually chosen to be
control performance.
Q(s)=F(s)GI’(s). F(s) is chosen to be a lowpass filter
with F(O)= 1 and GI’(s) a stable inverse of G(s). The
resulting controller takes the. following form 14 Magnitude optimum multiple integration
method
F(s)GI1(s)
C(S) =
1 -F(S)&+(S) The magnitude optimum multiple integration method
(MOMI) is based on the magnitude optimum (MO or
Therefore, a PI controller will result when a first-order BO) method [37] in which the frequency response from
system is to be controlled, and a PID controller when a
second-order system is to be controlled.
A special case of IMC is the 2-tuning method [30, 311,
which was developed for processes with long dead time
and is mainly described by first-order dynamics. Here the
controller transfer function is simply expressed as I Ho I
I
1 1 I
G,(s) = ~ - (43)
Gp(s)3,s
r
-
where Gc(s), Gp(s)and /z represent the controller’s transfer RO
function, the process transfer function and the desired
closed-loop time constant.
The overall closed-loop transfer function will take the
following form Fig. 5 K-B parametvised 2DF system
1
=-
1 As + (44)
Now using (43) and a P, PI or PID controller, the tuning
parameters can be obtained as a function of 1. Examples
related to ,?-tuning are presented in [30].
It can also be shown that if the system’s model is
accurate then by using 2-tuning the sensitivity is always
A
less than two. Small 1 gives a small IAE but increases the
system sensitivity. For further details of the ),-method the
reader is referred to [4, 30, 311.
In a similar approach, the zeros of the PID controller are
used to cancel out the system’s dominant poles. Although
this may work, especially when large time delays are
r
- S

Fig. 6 K-B parametvised opened system with augmented noise model

IEE Proc -Control Theory Appl.. Vol 149, No 1, January 2002 51
set-point to output should be as close to one as possible. If where G(s) is the transfer function of the plant and C(s) is
the closed-loop transfer function is GCL(s),then the objec- the transfer function of the controller used to shape the
tive is to find a controller such that plant's response.
The performance characteristics can be described as
GCL(0) = 1 bounds of the above functions in the frequency domain.
~ l G ; $ ~ ) l ~ u y o =0 (45) For good disturbance rejection and set-point following, the
sensitivity function S must be small for low frequencies.
For good noise attenuation and for robustness, related to
for as many n as possible [38]. This will give a fast non- high-frequency unmodelled dynamics, one wants the
oscillatory response for many cases. The MOM1 method complementary sensitivity T to be small.
can be used to tune a PID controller with transfer fbnction Resent research has been concentrated on the develop-
given by ment of techniques that combine identification and control
within the loop shaping framework for the design of more
reliable systems.
The frequency response information of the plant can be
employed for identification purposes to give its nominal
in the following way. First, the parameters Kp, b l , . . . , b,, model, together with estimated confidence limits. Convex
a1 ,. . . , a,, Tdelin the plant's transfer function model optimisation techniques, like I , or 12, can then be used to
tune the controller. The specifications are usually given in
the form of a desired loop transfer function (LTF).
In [41], a method that integrates identification and
tuning for PID control is presented. The objective of the
are obtained in the following way, using the concept of design is to maximise the disturbance attenuation. The
multiple integrations [39]. The process's open-loop step system to be identified is excited by a random binary signal,
response is recorded for a change AU and the following which is assumed to have sufficient energy content across
areas are calculated: the desired closed-loop bandwidth. The system's uncer-
tainty is estimated by considering a multiplicative uncer-
' 4 1 =v1(w) tainty structure. The selection of the target loop can then be
(48) done by optimising the disturbance rejection properties
subject to the uncertainty-related bandwidth constraints
with and the fundamental limitations on the plant's model
related to the locations of the plant's poles and zeros. It
is also pointed out that the target LTF must contain all non-
invertible dynamics present in the plant and compensator.
In the case where a more systematic approach is necessary,
one has to move from simple PID control to a full-order
(49) compensator design based on H, or other methods.
In the FLS PID tuning method, the PID coefficients K p ,
K , , Kd are tuned so that the compensated open-loop
Yk(t) =
JI
iAk-l -Yk-1(Z)1d7

Now, to meet (45) the following PID parameters are used

transfer function is close to the target transfer function
L(s) in a I , sense.
The PID controller transfer function is rearranged in the
following form:
[401
C(s) =
+
Kls2 K2s K3 +
s(7s 1)+ (55)

with K,, =K2 - K3s, K, =K 3 , Kd =K1 - K27 K3z2. +

The advantage of such a parameterisation, as stated in
[41], is that any functional of the form 11 W(GC - L)ll,,
with W a fixed weighting function, is convex in the new
design parameters.
(52) The FLS tuning of the PID controller can then be
formulated as the following optimisation problem:
This method is applicable to high-order systems, non-
minimum phase systems, and systems with reasonably
large time delays.
where So = (1 +L)-' is the target sensitivity and k is a
convex set of constraints for the PID parameters.
15 Frequency loop shaping FLS)
A more detailed description of this PID tuning method
and some applications of it can be found in [41, 421.
Frequency loop shaping methods are based on the use of
the sensitivity function S given by
16 Concluding remarks
1
S(S) =
1 + C(s)G(s)
(53)
This paper has presented an overview of PID control, its
and the complementary sensitivity function T given by advantages, disadvantages and different tuning methods.
Only a flavour of the available PID tuning methods has
T(s) = 1 - S(S) (54) been given, since it is not possible to include all available
52 IEE Proc.-Control Theory Appl., Vol. 149, No. 1, January 2002
methods here. Novel approaches to intelligent PID control,
using concepts of normalised dead time and normalised
process gain, are described in [ 5 ] . The symmetrical opti-
mum (SO) method [43] is also an optimisation method of
great interest. However, it must also be pointed out that
PID control may not be sufficient for some cases, for
example, processes with more than one oscillatory mode
or processes with large time delays or with complex
disturbance behaviour.
It is concluded here that PID control is still of great
interest, and is a promising control strategy that deserves
further research and investigation. Both industry and
academia have a lot to gain from this.
17 Acknowledgments
The authors would like to thank EPSRC and TRW for
supporting this research.
18 References
1 ASTROM, K.J., and HAGGLUND, T.: ‘Automatic tuning of PID
controllers’ (ISA, 1988)
2 ASTROM, K.J., and HAGGLUND, T.: ‘PID control’ in ‘The control
handbook’, 1996, pp. 198-209
3 ZIEGLER, J.G., and NICHOLS, N.B.: ‘Optimum settings for automatic
controllers’, Trans. ASME 1942, 64, pp. 759-768
4 HAGGLUND, T., and ASTROM, K.J.: ‘Automatic tuning of PID
controllers’ in ‘The control handbook’, 1996, pp. 817-826
5 ASTROM, K.J., HANG, C.C., PERSSON, I?, and HO, WK.: ‘Towards
intelligent PID control’, Automatica, 1992, 28, (I), pp. 1-9
6 SOYLEMEZ, M.T.: ‘Pole assignment for uncertain systems’ (UMIST,
Control Systems Centre Series, 1999)
7 ASTROM, K.J., and HAGGLUND, T.: ‘Automatic tuning of simple
regulators with specifications on phase and amplitude margins’, Auto-
matica, 1984, 20, pp. 645-65 1
8 HO, W.K., HANG, C.C,, and CAO, L.S.: ‘Tuning of PID controllers
based on gain and phase margin specifications’, Automatica, 1995, 31,
(3), pp. 497-502
9 SHENTON, A.T., and SHAFIEI, 2.: ‘Relative stability for control
systems with adjustable parameters’, 1 Guid. Control Dyn., 1994, 17,
pp. 304-310
10 SHAFIEI, Z., and SHENTON, A.T.: ‘Tuning of PID-type controllers for
stable and unstable systems with time delay’, Automatica, 1994,30, (lo),
pp. 1609-1615
11 SHAFIEI, Z., and SHENTON, A.T.: ‘Frequency-domain design of PID
controllers for stable and unstable systems with time delay’, Automatica,
1997,33, (12), pp. 2223-2232
12 BHATTACHARYYA, S.P., CHAPELLAT, H., and KEEL, L.H.: ‘Robust
control: the parametric approach’ (Prentice Hall, 1995)
13 CHAPELLAT, H., and BHATTACHARYYA, S.P.: ‘A generalization of
Kharitonov’s theorem: robust stability of internal plants’, IEEE Trans.
Autom. Control, 1989, 34, pp. 306-311
14 BARTLETT, A.C., HOLLOT, C Y , and LIN, H.: ‘Root location of an
entire polytope of polynomials: it suffices to check the edges’, Math.
Control Signals Syst., 1998, 1, pp. 61-71
15 HO, M.T., DATTA, A,, and BHATTACHARYYA, S.P.: ‘Design of P, PI
and PID controllers for interval plants’. Proceedings of the American
Control Conference, Philadelphia, Pennsylvania, June 1998, pp. 2496-
2501
16 MING-TZU, H., DATTA, A,, and BHATTACHARYYA, S.P.:1-‘
izations of the Hermite-Biehler theorem’, Lin. Alg. Appl., 1999,
pp. 302-303, 135-153
17 BULUT, B.: ‘The systematic design of PID controllers’, MSc thesis,
CSC, UMIST, Manchester, UK, 1998
IEE Proc.-Control Theory AppL, El. 149, No. 1, January 2002
18 XU, H., DATTA, A., and BHATTACHARYYA, S.P.: ‘Computation of all
stabilizing PID gains for digital control systems’, IEEE Trans. Autom.
Control, 2001,46, (4), pp. 647-652
19 MUNRO, N.: ‘The systematic design of PID controllers for continuous
and discrete systems’. IEE Colloquium on Advances in computer-aided
control systems design, Birmingham, June 1999
20 MUNRO, N., and SOYLEMEZ, M.T.: ‘Fast calculation of stabilizing
PID controllers for uncertain systems’. IFAC Conference, ROCOND,
Prague, June 2000
21 SILVA, G.J., DATTA, A,, and BHATTACHARYYA, S.P.: ‘Stabilization
of time delay systems’. Proceedings of the American Control Confer-
ence, Chicago, Illinois, June 2000, pp. 963-970
22 O’MAHONY, T., DOWNING, C.J., and FATLA, K.: ‘Genetic algo-
rithms for PID parameter optimization: minimizing error criteria’. Cork
Institute of Technology, Cork, Ireland, pp. 148-153
23 LIN, F., BRANDT, R.D., and SAIKALIS, G.: ‘Self-tuning of PID
controllers by adaptive interaction’. Proceedings of the American
Control Conference, Chicago, Illinois, June 2000, pp. 3676-3681
24 BRANDT, R.D., and LIN, E: ‘Supervised leaming in neural networks
without explicit error back-propagation’. Proceedings of the 32nd
Annual Allerton Conference on Communication, control and computing,
1994, pp. 294-303
25 LUENBERGER, D.G.: ‘Optimization by vector space methods’ (John
Wiley, 1969)
26 ISAKSSON, A.J., and GRAEBE, S.F.: ‘Model reduction for PID
design’. IFAC World Congress, Sydney, Australia, 1993, Vol. 8,
pp. 257-262
27 ISAKSSON, A.J., and GRAEBE, S.F.: ‘Analytical PID parameter
expressions for higher order systems’, Automatica, 1999, 35, (6),
pp. 1121-1130
28 RIVERA, D.E., MORARI, M., and SKOGESTAD, S.: ‘Internal model
control. Part 4: PID controller design’, Int. Eng. Chem. Process. Des.
Dev., 1986,25, pp. 252-265
29 GARCIA, C.E., and MORARI, M.: ‘Internal model control. Part 2: A
unifying review and some new results’, Ind. Eng. Chem. Process. Des.
Dev., 1982,21, pp. 308-323
30 THOMASSON, EY.: ‘Controller tuning methods’ in SELL, N.J. (Ed.):
‘Process control fundamentals for the pulp and paper industry’ (Tappi
Press, 1995), pp. 2 15-274
31 DAHLIN, EB.: ‘Designing and tuning digital controllers’, Insti: Control
syst., 1968, pp. 42-73
32 KAHNE, S.: ‘Pole-zero cancellations in SISO linear feedback systems’,
IEEE Trans. Educ., 1990, 33, (3), pp. 240-243
33 ISAKSSON, A.J., and GRAEBE, S.F.: ‘Model reduction for design of
digital PID controllers’. Proceedings of the 3rd European Control
Conference, Rome, Italy, September 1995, pp. 2191-2196
34 RIVERA, D.E., and MORARI, M.: ‘Control-relevant model reduction
problems for SISO H2, H,, and p-controller synthesis’, Int. 1 Control,
1987,46, (2), pp. 505-527
35 ANDERSON, B.D.O., and LIU, Y.: ‘Controller reduction: concepts and
approaches’, IEEE Trans. Aucom. Control, 1989, 34, (X), pp. 802-812
36 KEVICZKY, L., and BANYASZ, C.: ‘Designing PID regulators using
K-B-parametrization’. Process control and instrumentation. Glasgow,
2000, pp. 356-363
37 UMLAND, J., and SAFIUDDIN, M.: ‘Magnitude and symmetric opti-
mum criterion for the design of linear control systems-What is it and
how does it compare with the others?’, IEEE Ind. Appl. Soc. Trans.,
May/June 1990
38 ASTROM, K.J., and HAGGLUND, T.: ‘PID controllers: theory, design,
and tuning’ (Instrument Society o f America, 1995, 2nd edn.)
39 RAKE, H.: ‘Identification: transient and frequency response methods’
(System-s and Control Encyclopedia, August J 987)
40 VRANCIC, D., PENG, Y., and PETROVCIC, J.: ‘A new simple auto-
tuning method for PID controllers’. Preprints of the 2nd IFAC Workshop
on the New trends in design of control systems, Smolenice, August
1991, pp. 4 5 1 4 6 2
41 GRASSI, E., TSAKALIS, K., DASH, S., GAIKWAD, S.V,
MACARTHUR, W., and STEIN, G.: ‘Integrated system identification
and PID controller tuning bv freauencv looo-shanine’.
1 a , IEEE Trans.
Control Syst. Technol., 20;01,9, (2): pp. ?!85-i94
42 GRASSI, E., and TSAKALIS, K.: ‘PID controller tuning by frequency
loop-shaping: Application to diffusion fumace tempeGt&e control’,
IEEE Trans. Control Syst. Technol., 2000, 8, (5), pp. 842-847
43 VODA, A,, and LANDAU, LD.: ‘A method for the auto-calibration of
PlD controllers’, Automatica, 1995, 31, (l), pp. 41-53
53