Вы находитесь на странице: 1из 14

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/230872553

The future of PID control

Article in Control Engineering Practice November 2001


DOI: 10.1016/S0967-0661(01)00062-4

CITATIONS READS

574 600

2 authors:

K.J. strm Tore Hgglund


Lund University Lund University
310 PUBLICATIONS 25,896 CITATIONS 109 PUBLICATIONS 6,533 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Tore Hgglund on 24 April 2017.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
Control Engineering Practice 9 (2001) 11631175

The future of PID control


( stro. m*, T. H.agglund
K.J. A
Department of Automatic Control, Lund University of Technology, Box 118, S-22100, Lund, Sweden
Received 6 April 2001; accepted 6 April 2001

Abstract

This paper presents the state of the art of PID control and reects on its future. Particular issues discussed include specications,
stability, design, applications, and performance of PID control. The paper ends with a discussion of alternatives to PID and its
future. r 2001 Elsevier Science Ltd. All rights reserved.

Keywords: PID control; Stability; Performance; Application; Design; Tuning

1. Introduction Although the PID controller has always been very


important, practically it has only received moderate
Feedback is a very powerful idea. Its use has often interest from theoreticians. Therefore, many important
had revolutionary consequences with drastic improve- issues have not been well documented in the literature. A
ments in performance, see e.g. Bennett, (1979, 1993). result of this is that many mistakes have been repeated
Credit is often given to a particular form of feedback when technology shifted from pneumatic, via electrical
although it is frequently feedback itself that gives the to digital. There has, however, been an increased interest
real benets and the particular form of feedback used is in the last ten years. One reason is the emergence of
largely irrelevant. The PID controller is by far the most automatic tuning, another is the increased use of model
dominating form of feedback in use today. More than predictive control which requires well tuned PID
90% of all control loops are PID. Most loops are in fact controllers at the basic level. Still most papers on single
PI because derivative action is not used very often. loop control use PID controllers with ZieglerNichols
Integral, proportional and derivative feedback is based tuning as a benchmark. This is a very unsatisfactory
on the past (I), present (P) and future (D) control error. situation because the ZieglerNichols rules are known to
It is surprising how much can be achieved with such a give very poor results in many cases.
simple strategy. A strength of the PID controller is that This paper treats the future of PID control. It will try
it also deals with important practical issues such as to answer questions like: Will the PID controller
actuator saturation and integrator windup. The PID continue to be used or will it be replaced by other
controller is thus the bread and butter of automatic forms of feedback? What additional features are
control. It is the rst solution that should be tried when desirable in a PID controller? Are there any essential
feedback is used. research issues in PID control? Our prediction will be
The PID controller is used for a wide range of based on a consideration of the current state of research
problems: process control, motor drives, magnetic and and practice in PID control.
optic memories, automotive, ight control, instrumenta-
tion, etc. The controller comes in many dierent forms, 2. Specications
as standard single-loop controllers, as a software
component in programmable logic controllers and Before going into details it is important to be aware
distributed control systems, as a built in controller in that there is a wide range of control problems. A few
robots and CD players. examples are given below.

*Corresponding author. Tel:.00-46-461-08-781; fax: 00-46-461-38-


* Design a robust controller that keeps process vari-
118. ables reasonably close to desired values without
( stro. m).
E-mail address: karl-johan.astrom@control.lth.se (K.J. A strong demands on specications.

0967-0661/01/$ - see front matter r 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 9 6 7 - 0 6 6 1 ( 0 1 ) 0 0 0 6 2 - 4
1164 K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175

* Design a controller that keeps process variables as Notice that a real PID controller uses a ltered
close as possible to desired specications. derivative dyf =dt instead of dy=dt where
* Design a controller where the process variables can Td dyf
follow variations in set points. yf y:
N dt
* Design a controller that keeps process variables
within a range. Additional lters may also be used. The controller
should make sure that the integral part does not wind up
The last situation is typical for level control in surge when the actuator saturates. This is discussed in great
tanks where it is desired that the level changes but it is detail in A ( stro. m and H.agglund (1995).
not permitted either to have the tanks ooded or to have It is essential that the controller is implemented as (2)
them empty. with one integrator only. Parameters b and c are called
Another factor that has a strong inuence is the eort set-point weights. They have no inuence on the
that can be devoted to design and tuning of a system. response to disturbances but they have a signicant
One extreme case is found in process control where one inuence on the response to set-point changes. Set-point
process engineer may be responsible for several hundred weighting is a simple way to obtain a structure with two
loops. In such a case it is not possible to devote much degrees of freedom (Horowitz, 1963). It is also worth-
eort to each loop. Simplicity of handling and robust- while to observe that the so-called PI-PD controller
ness are then primary requirements. Another case is a (Atherton, 1999) is equivalent to set-point weighting.
dedicated system that is manufactured in large quan- See Taguchi and Araki (2000).
tities, for example, a feedback loop in a CD player. In Three external signals act on the control loop, namely
this case it is possible to devote a substantial eort to set point ysp ; load disturbance l and measurement noise
design a single control loop. n: The load disturbance drives the process variables
away from their desired values and the measurement
2.1. Formalization noise corrupts the information obtained from the
sensors.
To describe a design problem the process, the The design objective is to determine the controller
environment, and the requirements on the control have parameters in Gc s and Gff s so that the system
to be characterized. A typical situation is illustrated in behaves well. This means that the eect of load
Fig. 1. The process is described as a linear system with disturbances should be reduced, that too much measure-
transfer function Gs: The controller is also linear with ment noise should not be fed into the system and that
two degrees of freedom. The transfer function Gc s the system should be robust towards moderate changes
describes the feedback from process output y to control in the process characteristics. Therefore, the specica-
signal u; and the transfer function Gff s describes the tion will express requirements on
feed forward from set point ysp to u: For PID control,
one typically has
* Load disturbance response,
* Measurement noise response,
ki * Set point response,
Gc s k kd s;
s * Robustness to model uncertainties.
ki
Gff s bk ckd s: 1 The relations between the external input signals, ysp ; l
s and n; process output x measured signal y and control
which means that the inputoutput relation for the signal u are
controller can be described as G GGc GGff
x l n ysp ;
ut kbysp t  yt 1 GGc 1 GGc 1 GGc
Z t
ki ysp t  yt dt G 1 GGff
y l n ysp ;
0  1 GGc 1 GGc 1 GGc
dysp t dyt
kd c  ; 2 GGc Gc Gff
dt dt u  l n ysp 3
1 GGc 1 GGc 1 GGc

2.2. Criteria

Regulation performance is often of primary impor-


tance since most PID controllers operate as regulators.
Fig. 1. A block diagram describing a typical control problem. Regulation performance is often expressed in terms of
K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175 1165

the control error obtained for certain disturbances. A this section, it will be shown that the results can be
disturbance is typically applied at the process input. obtained from the Nyquist stability theorem. An
Typical criteria are to minimize a loss function of the advantage with this approach is that time delays are
form easy to deal with.
Z N
I tn jetjm dt; 4 3.1. Constant derivative gain
0

where the error is dened as e ysp  y: Common cases Consider linear systems with the transfer function
are IAE (n 0; m 1) or ITSE (n 1; m 2). Gs: It is assumed that the transfer function has no
Quadratic criteria are particularly popular since they poles in the right half-plane apart from possibly a pole
admit analytical solutions. Criteria which put a penalty at the origin. Furthermore, it is assumed that G0 > 0 if
on the control signal such as the system is stable or that lims-0 sGs > 0 when the
Z N
system has a pole at the origin. It is also assumed that
I e2 t ru2 t dt 5 the controller is parameterized in terms of k; ki and kd :
0
The closed-loop characteristic equation is then
have also been used.  
ki
1 k kd s Gs 0: 7
2.3. Robustness s
By analyzing this equation for small s it is seen that
It turns out that it is not sucient to minimize criteria ki > 0 to have a stable system. Next, conditions for roots
such as (4) or (5) because the solutions obtained may not on the imaginary axis will be investigated. This will
be suciently robust to uncertainties in the process represent a stability boundary. Introducing
model. Excellent insight into the robustness problem has Gio roeifo ;
been provided by the HN theory, see Panagopoulos and
Astro. m (2000). According to this theory it follows that after some calculations gives the following boundary of
the closed-loop system is robust to perturbations if the the stability region:
HN -norm of the transfer function cosfo
! k  ;
1 1 G ro
;
1 GGc Gc GGc o sinfo
ki o2 kd  : 8
ro
is suciently small. The norm is given by
1 jGioGc ioj Fig. 2 shows the stability region for a system with the
g max : 6 transfer function Gs 1=s 14 for dierent values of
j1 Gc ioGc ioj
kd : The gure shows that in this particular case the
In Panagopoulos and A ( stro. m (2000) it is shown that the integral gain can be increased substantially by using
norm is always less than M if the Nyquist curve of the derivative action. The gure also shows that the stable
loop transfer function of GGc is outside a circle with
diameter on the interval M=M  1; M=M 1:

3. Stability regions

Instability is the disadvantage of feedback. When


using feedback there is always a risk that the closed-loop
system will become unstable. Stability is therefore a
primary requirement on a feedback system. It turns out
that much insight into PID control can be obtained by
analyzing the stability region, which is the set of
controller parameters that give stable closed-loop
systems. This is also an area where there are interesting
recent results, see Shaei and Shenton (1994), Shenton
and Shaei (1994) and Anon (1999). These results are
based either on the HermiteBieler Theorem, see Ho,
Datta and Bhattacharyya (1996, 1997), or on elaborate
polynomial calculations, see Munro, So. ylemez and Fig. 2. Stability regions for Gs 1=s 14 for kd 0 (dashed), 5,
Baki (1999) and So. ylemez, Munro and Baki (1999). In 10, 15 and 20 (dotted).
1166 K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175

controller that maximizes the integral gain has para- constant derivative gain. Additional insight can be
meters which lie in a sharp corner. For systems with obtained from another representation of the stability
monotone transfer functions the stability region is a regions. To investigate the stability, the Nyquist
convex set but for other transfer functions the stability criterion is used and the locus of the loop transfer
region may consist of several disjoint sets. function Lio GioGc io is plotted. First a xed
The robustness region lies inside the stability region value k > 0 of the proportional gain is picked and the
and has a similar shape. This state of aairs is one frequencies on ; where the Nyquist curve of kGio
explanation why it is dicult to nd appropriate values intersects the circle with the line segment 1; 0 as a
of the derivative action. Other constraints can be diameter are determined. See Fig. 4, where the critical
introduced to deal with the problem, see Panagopoulos point s 1 is denoted by C. First, the case when the
(2000). Another possibility is to postulate a given value intersection of the Nyquist curve and the circle occur in
of the ratio of integral time to derivative time. the lower half-plane is considered. The controller
Classically it was quite common to postulate Ti 4Td : transfer function can be written as
Postulating that Ti mTd it is found that k2 ki kd =m:    
ki ki
Inserting this in Eq. (8) gives Gc io k i  kd o k  i  kd o :
o o
cosfo
k  ; When proportional and derivative gains are changed
ro
the Nyquist curve moves from A along the line AC. To
p avoid reaching the critical point it must be required that
1 4=m  1 cos2 fo  o sinfo  
ki 9 ki
2ro  kd on jGion joj1 Gion j: 10
on
Fig. 3 shows the stability regions for the system Gs
1=s 14 and PID controllers with m 4; 6.25 and The same analysis can be made when the intersection of
10. The gure shows clearly that the stability region may the Nyquist curve and the circle occur in the upper half-
be increased by using derivative action. But it also shows plane. To have a stable system the point A must then be
that the stability region becomes quite sharp when the moved beyond the critical point. Combining the
parameter m is small. A comparison with Fig. 2 shows conditions give the stability regions
that the largest value of integral gain ki is substantially ki > 0;
reduced by the requirement that Ti 4Td : j1 kGion j
ki oo2n kd on for Im Gion o0;
jGi on j
3.2. Constant proportional gain j1 kGion j
ki > o2n kd  on for Im Gion > 0 11
jGion j
The stability region is a subset of R3 : The calculations
performed gives the two-dimensional intersections with for all on such that
 
kGion 1 1: 12
2 2

It is thus concluded that for constant proportional gain,


the stability region is represented by several convex
polygons in the ki kd plane. In general, there may be

Fig. 3. Stability regions for Gs 1=s 14 and controllers with


Td 0 (dotted), and Ti mTd ; where m 10 (dashdotted), 6.25
(dashed) and 4 (solid). The stability region for pure PI control
corresponds to the dotted curve. Fig. 4. Nyquist curve for the transfer function kGs:
K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175 1167

several polygons and each may have many surfaces. The then stability requires that kd 0: If lim sG s-N
number of surfaces of the polygons is determined by the Kv then the integral gain is limited to jkd jo1=Kv :
number of roots of Eq. (12). In many cases, the
polygons are also very simple, as is illustrated with the
following example. 4. Design and tuning

Example 1 (Four equal poles). Consider a system with Design and tuning of PID controllers have been a
the transfer function large research area ever since Ziegler and Nichols
1 1 presented their methods in 1942. There are many aspects
Gs 4 that should be taken into account when designing a PID
s 14 s 4s3 6s2 4s 1
controller. Desirable features of a design procedure are:
1
4 : * It should give a controller that meets the design
s 6s 1 4ss2 1
2
specications.
In this case Eq. (12) becomes * It should be based on the available/obtainable
o4  6o2 1 k 0: process knowledge.
* It should meet limitations on computational power
This equation has only two positive solutions
p and resources available for design.
o2 37 8  k Therefore, there is a need for several dierent design
and it is thus concluded that the stability region is procedures with varying objectives and complexity.
determined by the condition ki > 0 and two lines. Simple Some recent results are Zhuang and Atherton (1993),
calculations show that the stability conditions (11) Lieslehto (1999), Anon (1999), Yamamoto, Fujii and
become Kaneda (1999), Panagopoulos (2000) and Wallen
p p (2000).
ki o3  8  kkd 4k  56 20 8  k;
p p
ki > 3 8  kkd 4k  56  20 8  k; 4.1. Specications for design
ki > 0: 13
The design specications were discussed in Section 2.
The integral gain has its maximum ki 36 at the Basically, one has to take care of specications on
boundary of the stability region for k 8 and kd 20: responses to the three external signals ysp ; l; and n;
Notice that the stability region is needle shaped near this as well as robustness with respect to changes in the
point. This is more clear in the 3D plot of the stability process G:
region shown in Fig. 5. Since the controller has two degrees of freedom, it is
possible to separate the specications. The response to
Analyzing systems with time delay it follows from the measurement noise can be treated by designing a low-
equations that if Gs approaches a constant for large s pass lter for the measurement signal. The desired set-
point response can be obtained by feedforward. One
way is to choose a proper value of the set-point weight b
in Eq. (2). It is also possible to feed the set point through
lters or ramping modules.
Load disturbances are often the most common
disturbances in process control. See Shinskey (1996).
Most design methods should therefore focus on load
disturbances, and try to nd a suitable compromise
between demands on performance at load disturbances
and robustness. It is a great advantage if this compro-
mise can be decided by the user by a tuning parameter.
There are two types of tuning parameters, those where
the robustness is specied, and those where the
performance is specied. Gain and phase margins are
often used as robustness-related design parameters. See
Tan, Wang, Hang, and H.agglund (1999). In A ( stro. m,
Panagopoulos, and H.agglund (1998) and Panagopou-
los, A( stro. m and H.agglund (1999), the maximum
Fig. 5. The 3D plot of the stability region for the system Gs sensitivity and the complementary sensitivity are used.
1 s4 : In Panagopoulos and A ( stro. m (2000) the HN norm is
1168 K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175

used. A common performance-related tuning parameter to obtain good tuning rules that are based only on two
is the apparent closed-loop time constant. It is used in process parameters, but that fairly accurate methods
lambda-tuning and in many other analytical tuning may be obtained using three parameters, where the
methods. relative gain k may be a third parameter.
The design procedure also includes selection of
controller structure. For example, if there are no 4.2. Process knowledge
requirements on zero steady-state error, there is no
need for integral action. If the dead-time is long or if the The process knowledge required by design procedures
process dynamics is close to rst order, no derivative varies from simple characteristics described by two
action should be used. parameters (a gain and a time) as in the ZieglerNichols
The design procedure should also take limitations in rules, to higher order transfer functions.
signals into account. There are always limitation in the In process control, there is normally little time
control signal, and often also in the rate of the control available to derive detailed models of the processes.
signal. The early design methods were therefore based on very
( stro. m et al. 1998 it is shown that PI control can
In A simple models. Nowadays, there are often equipments
be very well captured so as to maximize integral gain ki ; available that makes it possible to obtain the models fast
which is equivalent to minimizing the integrated error IE and simple. In automatic tuning procedures, the process
at load disturbances, subject to a robustness constraint. models are obtained fast with very little eort from the
The reason is that the parameters k and ki that satisfy user.
the robustness constraint is a good convex set. This does Many design methods assume that the process is
not work very well for PID control because the set of known completely in terms of step responses, frequency
parameters k; ki and kd that satisfy the robustness responses or transfer functions. Actually, it is sucient
constraint can have narrow ridges. See Panagopoulos to know the frequency response in a fairly narrow
(2000). This is illustrated in Fig. 5 which shows the frequency range. For the design method in Panagopou-
stability region for a problem that appears very well los (2000) the transfer function should be known in a
behaved. region around opi* and opd* ; where
The design method given in A ( stro. m et al. 1998 can
also be used to evaluate simpler design methods. Fig. 6 1
arg Giopi*  p arcsin ;
shows controller parameters obtained for a large test Ms
batch of process models. The gure also shows the
ZieglerNichols design. The controller gain is normal- *  p  arccos 1
arg Giopd :
ized with the ultimate gain, and the integral time is Ms
normalized with the ultimate period of the process. The
parameters are plotted versus the relative gain
jGio180 j 1 4.3. Computational aspects
k : 14
G0 Kp Ku
The equipment, the computational power, and
Results are given for two values of the design the time available for the design are also important
parameter Ms (the maximum sensitivity gain), Ms aspects that must be taken into account. The
1:4 and Ms 2: early methods were, for obvious reasons, based on
The gure shows that the ZieglerNichols method simple registrations of process characteristics and
gives too high a gain and for most processes an integral calculations that could easily be performed by hand.
time that is too long. It also shows that it is impossible There is still an interest in these methods for fast
controller settings. These methods are also useful for the
understanding and intuition needed for manual con-
troller adjustments.
Nowadays, there is often a computational power
available that admits more advanced process identica-
tion as well as controller design. In automatic tuning
procedures, the identication and design are both made
automatically. The process models used in these
procedures are normally still rather simple. The design
procedures are, however, often much more sophisti-
Fig. 6. Normalized controller parameters plotted versus relative gain k cated. It is, for example, possible to use optimization
for Ms 1:4 (circles) and Ms 2 (crosses). The ZieglerNichols rule is routines to obtain optimal solutions stated by the design
shown in dashed lines criteria given in Section 2.
K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175 1169

5. Performance assessment Process uncertainty is used explicitly in design


methods such as HN (Zhou, Doyle, & Glover, 1996)
It is useful to have tools to make an assessment of the and QFT (Horowitz, 1993). Some consequences of this
performance that can be achieved and the factors that for PID control are discussed in Panagopoulos and
limit the performance. This is a research topic that has o. stro. m (2000) and Fransson, Lennartsson, Wik, and
recently attracted much attention, see Seron, Braslavsky Gutman (2000).
and Goodwin, (1997) and A ( stro. m (2000).
For a control system the achievable performance is 5.3. When is derivative action useful?
typically limited by
In Section 3, it was found that adding derivative
* Process dynamics.
action to a PI controller increases the complexity of the
* Nonlinearities.
design considerably. There is much folklore concerning
* Uncertainties.
derivative action. Why is the derivative term hard to
* Disturbances.
tune? Why should one choose Ti 4Td ? What is the use
It is essential to be aware of the factors that are crucial of complex zeros, i.e. Ti o4Td ? Why is the derivative
for a specic application. Most of the conventional action not suitable for a pure dead-time process. To get
design methods for PID control focus on process some insight into these problems, the results on
dynamics and uncertainties. performance assessment will be used. First, a specic
example is discussed.
5.1. A preliminary assessment
Example 2 (Process with lag and delay.). Consider a
To make a preliminary assessment, the gain crossover process with a rst-order lag and a time delay, i.e.
frequency ogc will be used to characterize performance. 1
Let jm be the desired phase margin, the crossover Gs esL : 16
1 sT
frequency is the smallest frequency such that Conditions for when the process can be successfully
argGp iogc argGc iogc p jm : 15 controlled by a PI controller will rst be determined. It
follows from the discussion above that the crossover
A PD controller has a maximum phase lead of about
frequency is limited by ogc : It follows from Eq. (15) that
60 degrees, a proportional controller has zero phase
lag, and a PI controller has a phase-lag of about arctanogc T ogc L  argGc io p  jm :
45 degrees and a PID controller can have a phase lead Assume that a phase margin of 60 degrees is desired
of 45 degrees. If a phase margin of 45 degrees is desired it and that PI control is used. Since ogc Lo1 it follows that
follows from Eq. (15) that crossover frequencies for PI, arctanogc To2p=3  p=4  1 0:3090: It can thus be
PID and PD control are the frequencies where the
process has phase-lags of 90 degrees, 135 degrees and
195 degrees, respectively. These frequencies are denoted
as o90 ; o135 ; and o195 : These estimates are valid only if
the non minimum phase parts of the dynamics are not
dominating.
( stro. m (2000) it is shown that for minimum phase
In A
systems the phase lag of the minimum phase parts has to
be less than 60 degrees at the crossover frequency. The
system will have poor robustness if the phase lag is
larger. For a system with dead time this implies that the
crossover frequency is approximately limited to
ogc Lo1: This result can be used to determine when
the PI control can be used for processes with time delay.

5.2. Process uncertainties

Process uncertainties impose severe limitations on


control performance. Traditionally, this has been taken
into account using gain and phase margins. The design
Fig. 7. Responses in process output to steps in load disturbances for
methods developed by Kessler for motor drives used the PD, PI and PID controllers for systems with the transfer function
slowest neglected time constant as an important part of Gs Tes =1 sT with (a) T 10; (b) T 3 (c) T 1 and (d)
the specications. See Kessler (1958a,b). T 0:3:
1170 K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175

concluded that PI control can be used if To0:3L or if tive action. This shows clearly how important it is to
L > 3T: react fast in order to achieve good control. Also notice
Fig. 7 shows responses to load disturbances for PD, that at a casual inspection the dierences in the control
PI and PID controllers for four dierent systems with signal do not appear very large but that they have a
dierent ratios of dead-time to time constant. Notice profound eect on the output signal.
that derivative action gives substantial benets for
systems having large values of T: Due to the predictive The examples support the folklore that derivative
action the error starts to decrease much faster. Notice action is useful for processes that are lag dominated and
that the peak disturbance obtained for PID control is that it is less useful for systems that are dead-time
lower than for PD control. dominated, see Shinskey, (1996).
All controllers simulated in Fig. 7 are designed by
maximizing integral gain subject to the robustness
constraint Ms o 2 using the method described in 6. Competing strategies
Panagopoulos (2000), where the constraint Ti 4Td
was imposed on the PID controller, see Wallen, (2000). This section will discuss some situations where there
The gure shows that there is a signicant improve- are competing strategies that can do better than PID
ment by using derivative action for small values of L=T control. Four dierent cases will be discussed dead-time
but that the benet of derivative action decreases when dominant systems, oscillatory systems, multivariable
the ration L=T increases. systems, and nonlinear systems. It has attempted to
answer the question when it is possible to obtain drastic
To get more insight into the question, when derivative improvements over PID control.
action is useful the extreme case of an integrator with
delay will be considered.
6.1. General linear controllers
Example 3 (Integrator with delay.). Consider a system
The PID controller with only a few parameters is
with the transfer function
certainly a restricted complexity controller. An alter-
1 native would be to replace it by a general linear
Gs es : controller. A general two degree controller can be
s
represented as
For this system o90 0; o135 0:78; and o195 1:8:
These values indicate that the crossover frequency can RsUs TsYsp s  SsYs;
be doubled by using integral action. To get more insight
where R; S and T are polynomials of arbitrary order.
PD, PI and PID controllers were designed to maximize
Another alternative would be to use a state space
ki subject to Ms o2 using the method described in
representation consisting of an observer with state
Panagopoulos (2000), where the constraint Ti 4Td
was imposed on the PID controller, see Wallen (2000).
The controller parameters obtained are given in Table 1.
Notice that the integrated error IE is reduced from 7.6
for PI control to 2.2 for PID control and that the
integrated absolute error IAE is reduced from 9.2 for PI
control to 3.8 for PID control.
Fig. 8 shows the responses to a step change in load
disturbance and the control signal for PD, PI and PID
controllers designed to maximize ki subject to Ms o2:
The gure shows that the major benet in reducing the
eect of the load disturbance is that the control signal
reacts much faster initially for controllers with deriva-

Table 1
Parameters of PD, PI and PID controllers for a system with the
transfer function Gs es =s:
Type k ki kd Ti Td

PD 0.854 0.255 0.299


Fig. 8. Responses in process output to a step in the load disturbance
PI 0.488 0.131 3.739
applied at time t=0 for PD (dashdotted), PI (dashed), and PID
PID 0.864 0.462 0.404 1.869 0.467
(solid) control of a system with the transfer function Gs es =s:
K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175 1171

feedback and model following. Such controllers are where Gsp is the Smith predictor given by
straight forward extensions to PID control. Some 1
problems with these general controllers are parameter- Gsp : 17
1 G* G* c 1  esL
ization, design and tuning. General linear controllers
include PID control as a special case and can thus If G* G* c E  1 then Gsp EesL which indicates that Gsp acts
as a predictor in a certain frequency range, see A ( stro. m
always outperform a PID controller. It is however a
nontrivial question to nd good man-machine interfaces (1977).
so that the users can easily specify structure and The transfer function from the set point to the process
parameters. Tuning tools are also necessary. There are output is given by
a number of commercial linear adaptive controllers but Ys G* G* esL
they are at present much less common than PID 18
Ysp s 1 G* G* c
controllers, see Bengtsson and Egardt (1984) and
Mod!en (1995). The transfer function from load disturbance to the
process output is
6.2. Set point response Ys 1 G* G* c 1  esL * sL
Ge : 19
Ls 1 G* G* c
It has been emphasized several times that the problem
of regulation and set point response should be A particularly simple case is when G* Gpi kv =s
separated, see Fig. 1. Our approach was to rst design 1=sT: To see that Gsp acts like a predictor, a series
the feedback Gc to obtain a system that will balance the expansion in s is made
 
eects of load disturbances and measurement noise and 1 kv L L
Gsp E 1 s :
which is robust to process uncertainties. The feedfor- 1 kv L 1 kv L 2
ward path Gff is then designed to obtain the desired
For large kv the predictor thus predicts over a time
response to set point changes. This results in the two-
horizon of L=2: Assuming that the process has a unit
degree of freedom structure advocated in Horowitz
gain at low frequencies, the eective integral gain of a PI
(1963). For set point changes of moderate size good
controller with a Smith predictor becomes
results can often be obtained by using set point
weighting, see A ( stro. m and H.agglund (1995). In some kv 1
ki : 20
cases it may be desirable to add linear lters to reduce 1 kv L T L
the overshoot, see Panagopoulos (2000). Rapid set point It follows from this equation that it is desirable to
responses are often limited by the nonlinearities of the have a large value of kv or small values of T: The largest
process. Recent work by Wallen (2000) indicates that value that can be used is limited by the requirement on
there are substantial benets by using nonlinear control robustness.
strategies. For systems with time delays the response The sensitivity function is
time can be decreased substantially by using Smith
predictors. 1 G* G* c 1  esL G* G* c esL
Gs 1 :
1 G* G* c 1 G* G* c
6.3. Processes with time delays For the particular design the sensitivity function
becomes
There is much folklore concerning systems with time
kv esL esL
delay. Such systems can be controlled quite well with Gs 1  1 :
PID control. However, the traditional tuning rules often s kv sT 1
give very poor results. Derivative action is quite useful The maximum values of the sensitivity and the
for systems which also has lags as was found in Section complementary sensitivity are thus less than 2 for all
5. Derivative action, however is of limited value for values of T: A sensitivity analysis thus indicates that
systems that are dead time dominant. The reason for arbitrary small values of T can be used if sensitivities
this is that prediction of the output based on the linear less than 2 are admitted. This however does not tell the
extrapolation is not eective. It is much better to make full story.
predictions based on inputs that were fed into the system Fig. 9 shows the Nyquist curves for the loop transfer
than those which have not yet shown up in the output. function for dierent values of the ratio T=L: Notice
The Smith predictor accomplishes this, see Kaya and that there are loops in the Nyquist curves that increase
Atherton (1999), Kristiansson and Lennartson (1999) with decreasing values of T: Because of these loops the
and Jiya, Shao and Chai, (1999). system will be sensitive to variations in the time delay.
Let the process transfer function be G G* esL : A Systems of this type are classic cases where the
controller with a Smith predictor can be interpreted as a sensitivity functions do not give good insight into the
controller of the type shown in Fig. 1 with Gc G* c Gsp ; robustness issues. It is necessary to also ensure that the
1172 K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175

Fig. 10. Responses to a step in load disturbance for a system with a


Smith predictor with kv 5 (dashed) and PI control with k 0:25 and
Ti 0:35L (solid).

Fig. 9. Nyquist curves for the loop transfer functions for T=L 0.1
and 0.2 for the nominal case (left) and for a system with a 25%
increase of the time delay (right).

6.4. Systems with oscillatory modes

Systems with oscillatory modes can be controlled


delay margins are satisfactory. Fig. 9 shows, for
with PI controllers if the requirements are not too
example, that a system with T=L 0:1 becomes
high. The performance of a PID controller can,
unstable if the time delay is increased by 25% from L
however, be improved by adding lters to the control
1 to 1:25: system.
It is of interest to compare the results that can be
obtained with simple PI control. This is made in the
6.5. Multivariable systems
following example.
Many processes are multivariable. Much research
Example 4 (A pure delay process). Consider a pure
has lately been devoted to control schemes for
delay process which has the transfer function G esL :
such systems, see e.g. Morari and Lee (1991), Camacho
The integral gain obtained with Smith predictor control
and Bordons (1995). Most multivariable schemes do
is given by Eq. (20). It gives ki 0:8=L for kv 5: however operate in cascade mode, where the multi-
Designing a PI controller to maximize ki subject to a
variable controller provides set points to local PID
robustness constraint Ms o2 gives k 0:25 and Ti
controllers, see Lee, Park and Lee, (1999). The PID
0:35L; hence ki 0:71=L: The Smith predictor thus
controller is thus an essential building block, also in this
gives only a very small increase of ki compared to the PI
case, tuning can often be done iteratively, see V!asquez,
controller. Fig. 10 compared the responses to load
Morilla and Dormido (1999). It should also be
disturbances of systems with PI control and with PI
emphasized that PID controllers with static decoupling
control and a Smith predictor. Notice that the control
can do very well for multivariable systems if the
signal for the Smith predictor is smoother. This is an requirements are not too high. This is illustrated by an
advantage because it means that modes of higher
example.
frequencies are not excited.
Example 5 (Rosenbrocks system). A seemingly simple
The results indicate that for systems with pure time
multivariable system proposed by Rosenbrock has the
delays the Smith predictor only gives marginal improve-
transfer function
ments in regulation performance. These improvements 0 1
may be relevant for control of important quality control 1 1
Bs 1 s 1C
problems but not for ordinary problems. The Smith Gs B
@ 2
C:
predictor may, however, give substantial improvements 1 A
in the response to set point changes. s3 s1
K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175 1173

By static decoupling a system with the transfer function eective and can be applied to a wide range of problems.
0 1 The emerging features of automatic tuning have greatly
1
B 0 C simplied the use of PID control.
s1
Qs GsG1 0 B
@
C; More knowledge about PID control has been avail-
4s 31  s A
able for a long time. Unfortunately, it has been buried in
s 1s 3 s 1s 3
proprietary information of suppliers. There was a strong
is obtained. This shows that the system has a right half- resurgence in the interest in PID control over the last 10
plane zero which explains why it is dicult to control. It years. Many publications have appeared and it is typical
is, however, possible to control this system by PID that IFAC organized a workshop on PID control in the
controllers provided that the demands are not excessive. year 2000.
This is illustrated in the step responses shown in Fig. 11. The alternatives to PID control are:
It is also possible to provide simple estimates for how
fast the system can be made while keeping the RST: Discrete-time linear MISO controllers.
interactions at a reasonable level. SFO: State feedback and observers.
MPC: Model predictive control.

6.6. Nonlinear systems Fuzzy control is often mentioned as an alternative to


PID control, see Passino and Yurkovich (1998). Most
The PID controller being linear is not suited for fuzzy controllers used in industry have the same
strongly nonlinear systems. Excellent results can how- structure as incremental PI or PID controllers. The
ever be obtained by combining the PID controller with parameterization using rules and fuzzy membership
gain scheduling. A particularly attractive feature is the functions makes it easy to add nonlinearities, logic, and
use of auto-tuning which drastically reduces the eort additional input signals to the control law.
required to build up the gain schedule. The main advantages claimed by fuzzy control are
that it is easy to use and that it is nonlinear. Many of the
comparisons between PID and fuzzy control made in the
literature are, however, very sloppy. A textbook PID
7. Does PID control have a future?
with ZieglerNichols tuning is typically used as a
reference. Furthermore, if nonlinear behavior is desired
It is quite reasonable to predict that PID control will
continue to be used in the future. Feedback has had a gain scheduling can be added to a PID controller. Many
fuzzy controllers are also used in a cascade structure
revolutionary inuence in practically all areas where it
using PID controllers at the lower level. An advantage
has been used and will continue to do so. PI(D) control
of fuzzy control is that very good software is available.
is perhaps the most basic form of feedback. It is very
Many fuzzy controllers are, however, used in a cascade
structure using PID controllers at the lower level.
To assess the dierent alternatives it is important to
consider.

* Performance.
* Tuning.
* Ease of use.
* Maintenance.

All alternatives oer potential improvements in the


linear behavior of the system. This is particularly useful
when dealing with systems having poorly damped
oscillatory modes. It is, however, necessary to take
actuator saturation and windup into account. This is
done very well for MPC but it requires special
consideration for RST and SFO, particularly in
combination with mode changes.
The tuning problem is a major diculty with all
alternatives. Much work is needed in order to develop
Fig. 11. Simulation of PI control with static decoupling of Rosen-
brocks system. The gure shows the response of the outputs to steps in appropriate tuning tools.
the command signals. All controllers have set point weighting with The RST controller is beginning to appear in
b 0: applications for motor drives. The improvements in
1174 K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175

performance are particularly important for high perfor- need for a wide range of tuning techniques. There
mance systems. are already many tuning methods available, but a
Controllers based on state feedback and observers are replacement of the ZieglerNichols method is long
used in special applications where the cost of the overdue. It is very easy to demonstrate that any
modeling eort can be justied. controller with reasonable tuning will outperform a
Model predictive control is typically used in super- PID with ZieglerNichols tuning. Many strategies
visory mode with PID controllers at the base level. proposed can easily be eliminated if they are compared
Much of the improvement accredited to MPC in the with a well-tuned PID.
process industry actually comes from the improved Development of suitable software is another area that
tuning of the basic loops. The MPC does, however, oer has to be developed. PID control is quite under-
drastic improvements in set point responses for multi- developed in that respect compared, for example, to
variable systems because of the coordination it provides. fuzzy control. It would be highly desirable to have
There are several interesting problems related to the software so that persons with a moderate knowledge can
integration of MPC and PID. Tuning of the PID experiment with PID control. Tools for modeling and
controller in the inner loop gives valuable modeling methods for automatic tuning should be a part of such a
information for the MPC. Since the MPC operates in a software.
supervisory mode it can deal with slow interactions very On the research side, it appears that the development
well. The lower level PID loops still have to manage fast of design methods for PID control is approaching the
interactions. point of diminishing returns. There are some dicult
Another issue is the general philosophy of approach- problems that remain to be solved. For example, there is
ing design of a complex system. Topdown approaches no characterization of the processes where PID control
to system design will clearly favor SFO and MPC. A is useful.
bottomup approach favors the use of simple building
blocks such as PID controllers. In this context, it is
interesting to see how the PID controller can be References
augmented. The set point response can be improved
substantially by exploiting a controller structure with Anon, A. (1999). Special edition on PID tuning methods. Computing &
two degrees of freedom as illustrated in Fig. 1 in this Control Engineering Journal, 10(2).
( stro. m, K. J. (1977). Frequency domain properties of Otto Smith
A
paper. Special considerations have to be given to a PID regulators. International Journal of Control, 26, 307314.
controller that will be used eectively in this way. It is ( stro. m, K. J. (2000). Limitations on control system performance.
A
particularly important that the controller output is European Journal on control.
( stro. m, K. J., & H.agglund, T. (1995). PID controllers: Theory, design
A
available for feedforward in such a way that saturation
and windup are handled properly. and tuning. Research Triangle Park, N.C. Instrument Society of
America.
More improvements can be made through proper use Atherton, D. (1999). PID controller tuning. Computing & Control
of feedforward. Other possibilities are to add lters and Engineering Journal, 4450.
blocks for dealing with measurement noise and systems Bengtsson, G., & Egardt, B. (1984). Experiences with self-tuning
having oscillatory modes. There are many good research control in the process industry. Preprints 9th IFAC world congress,
Budapest, Hungary (pp. XI: 132140).
problems in developing good approaches to bottom up
Bennett, S. (1979). A History of Control Engineering 18001930.
design techniques. The problem of exploring system London: Peter Peregrinus.
interactions is an important one. Bennett, S. (1993). A History of Control Engineering 19301955.
Even if the applications of other control strategies London: Peter Peregrinus.
increase, PID control will certainly continue to be used. Camacho, E. F., & Bordons, C. (1995). Model prediction control in the
When correctly used it is a very eective way of using process industry. Advances in industrial control. Berlin: Springer.
Fransson, C., Lennartsson, B., Wik, T., & Gutman, P. (2000). On
feedback. Good results can often be obtained if the optimizing PID controllers for uncertain plants using Horowitz
performance requirements are not extreme. The PID bounds. IFAC workshop on digital controlFpast, present, and
controller will also serve as a building block in more future of PID Control, Terrassa, Spain.
complex controllers. Most DMC controllers in fact Ho, M. T., Datta, A., & Bhattacharyya, S. P. (1996). A new
deliver set points to PID controllers. Good performance approach to feedback stabilization. In Proceedings of the
35th IEEE conference on decision and control, IEEE, vol. 4
of these PID controllers are essential. Much commis- (pp. 46434648).
sioning work for DMC control actually consists of Ho, M. T., Datta, A., & Bhattacharyya, S. P. (1997). A linear
tuning up the underlying PID loops. There are also programming characterization of all stabilizing PID controllers.
useful augmentations for PID control in the form of Proceedings of the American control conference, IEEE, Albuquer-
que, NM (pp. 39223928).
Smith predictors, gain scheduling and lters for
Horowitz, I. (1993). Quantitative feedback theory (QFT). Boulder, Co:
oscillatory systems. QFT Publications.
It is important to realize that there is a very wide Horowitz, I. M. (1963). Synthesis of feedback systems. New York:
range of control problems and consequently also a Academic Press.
K.J. Astrom, T. Hagglund / Control Engineering Practice 9 (2001) 11631175 1175

Jiya, J., Shao, C., & Chai, T. Y. (1999). Comparison of PID and PPI Panagopoulos, H., A ( stro. m, K. J., and H.agglund, T. (1998). Design of
design techniques for a process with time delay. Preprints. 14th PI controllers based on non-convex optimization. Automatica,
world congress of IFAC, Beijing, China (pp. 391396). 34(5) 585601.
Kaya, I., & Atherton, D. P. (1999). A new pipd smith predictor for Passino, K. M., & Yurkovich, S. (1998). Fuzzy control. Menlo Park,
control of processes with long time delays. Preprints. 14th world CA. Addison-Welsley-Longman.
congress of IFAC, Beijing, China (pp 283288). Seron, M. M., Braslavsky, J. H., Goodwin, G. C. (1997). Fundamental
Kessler, C. (1958a). Das symmetrische Optimum, Teil I. Regelung- limitations in ltering and control. Berlin: Springer.
stechnik, 6(11), 395400. Shaei, Z., & Shenton, A. T. (1994). Tuning of PID-type controllers
Kessler, C. (1958b). Das symmetrische, Optimum Teil II. Regelung- for stable and unstable systems with time delay. Automatica,
stechnik, 6(12), 432436. 30(10), 16091615.
Kristiansson, B., & Lennartson, B. (1999). Optimal PID controllers Shenton, A. T., Shaei, Z. (1994). Relative stability for control systems
including roll o and smith predictor structure. Preprints. 14th with adjustable parameters. Journal of Guidance Control and
world congress of IFAC, Beijing, China (pp. 297302). Dynamics, 17(2) 304310.
Lee, Y., Park, S., & Lee, J. H. (1999). On interfacing model predictive Shinskey, F. G. (1996). Process control systems. Application, design and
controllers with low-level loops. Preprints. 14th world congress of tuning (4th ed.) New York: McGraw-Hill.
IFAC, Beijing, China (pp. 313318). So. ylemez, M. T., Munro, N., Baki, H. (1999). Fast calculation
Lieslehto, J. (1999). Collection of java applets for PID controller of stabilizing PID controllers. Automatica, submitted for publi-
design. Preprints. 14th world congress of IFAC, Beijing, China (pp cation.
421426). Taguchi, H., & Araki, M. (2000). Two-degree-of-freedom PID
Mod!en, P. E. (1995). Experiences with adaptive control since 1982. controllerstheir functions and optimal tuning. IFAC workshop
Proceedings of the 34th IEEE conference on decision and control, on digital controlFPast, present and future of PID Control,
New Orleans, LA, pp. 667672. Terrassa, Spain.
Morari, M., & Lee, J. H. (1991). Model predictive control: The good, Tan, K. K., Wang, Q. G., Hang, C. C., & H.agglund, T. (1999).
the bad, and the ugly. Chemical process control, CPCIV, Padre Advances in PID control. Advances in industrial control. Berlin:
Island Tx (pp 419442). Springer.
Munro, N., So. ylemez, M. T., & Baki, H. (1999). Computation of D- V!asquez, F., Morilla, F., & Dormido, S. (1999). An iterative method
stabilizing low-order compensators. IEEE Trans. on Automatic for tuning decentralized PID controllers. Preprints. 14th world
Control, Submitted for publication. congress of IFAC, Beijing, China (pp. 491496).
Panagopoulos, H. (2000). PID control design, extension, application. Wallen, A. (2000). Tools for autonomous process control. Ph.D. thesis,
Ph.D. thesis, Department of Automatic Control, Lund Institute of Department of Automatic Control, Lund Institute of Technology,
Technology, Lund, Sweden. Lund, Sweden.
Panagopoulos, H., & A ( stro. m, K. J. (2000). PID control design and Yamamoto, T., Fujii, K., & Kaneda, M. (1999). A design of self-tuning
HN loop shaping, International Journal of Robust and Nonlinear PID controllers based on a pole-assignment scheme. Preprints. 14th
Control, 10, 12491261. world congress of IFAC, Beijing, China (pp. 259264).
Panagopoulos, H., & A ( stro. m, & H.aglund, T. (1999). Design of PID Zhou, J., Doyle, J., & Glover, K. (1996). Robust and optimal control.
controllers based on constrained optimization. Proceedings of the Englewood Clis, NJ: Prentice-Hall.
1999 American control conference (ACC99), San Diego, CA. Zhuang, M., & Atherton, D. P. (1993). Tuning of optimum PID
Invited paper. controllers. Proceedings of IEE, 140, 216224.

View publication stats