Professional Documents
Culture Documents
www.elsevier.com/locate/jprocont
Abstract
An extensive study of robust and optimal tuning of PID controllers for stable non-oscillating plants is presented. It is built on a
set of well dened criteria related to output performance, stability margins and control activity. Dierent interesting properties of
the closed loop systems are observed. A set of simple tuning rules is based on these observations. These rules are compared to a
couple of well established tuning methods and are shown to give well competitive results, especially when simplicity, low control
activity and high-frequency robustness are emphasized. Derivative action is shown to improve performance signicantly compared
to PI control, with equal stability margin and a moderate increase of control activity, for most plants, including those with significant time delay.
2005 Elsevier Ltd. All rights reserved.
Keywords: PID control; Robustness; Optimization; Optimal systems; Time delay
1. Introduction
An objective method to evaluate the trade-o between performance and robustness in dierent frequency
regions for all kinds of linear controllers has recently
been presented by the authors [13]. This evaluation
method is briey presented in Section 2.
PI and PID controllers for a large amount of plant
models with all their poles on the negative real axis have
been optimized for good rejection of load disturbances
according to the suggested evaluation method. This is
motivated by the fact that most PID controllers work
as regulators [4,5]. For most investigated plants the optimal PID controller has a pair of complex zeros, which
has lead to a reformulation of the traditional PID con-
troller. The new formulation is presented in Section 3 together with the benchmark models used.
Section 4 is devoted to a series of general properties
for optimal PID controlled systems. It is noticed that
very limited improvements in terms of performance
can be attained, even with considerable raise of the control activity beyond a certain level. Moreover, the bandwidth of the optimal system is typically just above the
phase crossover frequency for the plant, and a PID controller designed for good disturbance rejection will in
most cases introduce a signicant negative phase shift.
It is also observed that the design method used will
place the resulting poles very similar for all plants
investigated.
An extremely simple method for design of PID controllers is presented in Section 5. All plant knowledge
that is asked for can be found by a simple step response.
An extended description of this and some related tuning
methods is given in [6,7]. The presented method is compared to some common and well appreciated tuning
methods. It is shown that the new method well asserts
92
2. Evaluation method
The proposed performance criterion is mainly a measure of the systems ability to handle low-frequency load
disturbances, a frequency domain alternative to the
more commonly used criteria based on a function of
the error signal [9,10]. It is formulated as
1
1
J v Gyv S v s
1
s
s
1
1
When there is integral action in the controller, L(s) =
G(s)K(s) 1 and Sv(s) = G(s)/(1 + G(s)K(s)) K1(s)
s/ki for low-frequencies (ki is the integral gain). Thus, in
this frequency range Jv 1/ki has the advantage of
being almost independent of the plant model.
Alternative performance criteria: There are several different performance criteria proposed in the literature.
The most commonly used is presumably
Z 1
IAE
jetj dt
0
1
Ger s
1 Ls
S v s
Gs
Gyv s
1 Ls
Ks
Gur s Guw s
1 Ls
10
ki
1/J
|Sv|
v
r
+
-
K(s)
G(s)
Fig. 1. Closed loop SISO system with plant G(s) and controller K(s).
10
-1
10
93
This criterion converts the measure (3) to a corresponding MS level. A similar criterion was formulated
by Schei [19]. In [9] two suitable values for MS (1.4
and 2.0) are proposed. Throughout this paper the default values are MS = 1.7 and MT = 1.3, which means
that an optimal controller with GMS = 1.7 results in a
system with Gm P 2.4 and um P 45. For high-gain
controllers, such as Smith predictors, a demand on
bounded high-frequency loop-gain may also be included
in GMS, for more details see [2].
2.3. Control activity
especially for unstable plants including those with integral action, see [19,20,8,21].
1.2
max ki
0.2
1
0.8
0.3
min J
Design of a control system is typically a trade-o between performance and its price in terms of control
activity, as soon as demanded stability margin is guaranteed. Therefore, introduce the control activity criterion
min t
min ts
min IAE
0.1
min IAE
0.6
max b
0.4
min Jv
0.2
max
max k
0
0
10
0.1
Fig. 3. Reference (left) and load disturbance (right) step responses for G(s) = e
0.3s
10
94
J u q 6 C 2
7
sT i 1 sT f
Very often, ever since the days of Ziegler and Nichols,
the relation a = Ti/Td has been given a xed value a = 4
[2628], corresponding to a double zero in a controller
without derivative lter. In the same way b = Td/Tf
has been xed, often to 10, [2931].
3.1. Alternative tuning parameters
In e.g. [32,21] it has been shown that a PID controller, optimized with the lter included and all parameters
free, often implies complex zeros in the controller. This
has been veried for a large number of plants with poles
strictly on the negative real axis. Hence, a suitable alternative parameterization of the PID controller is
K PID s k i
1 2fss ss
s1 ss=b
Tf
s
b
T i 2fs T f
Td
s2
Tf
Ti
K p K iT i
jGjx150G j
G0
10
with the four tuning parameters the integral gain ki, the
high-frequency gain k1 = KPID(1) = kisb, the zero
damping f and the zero time constant s. The notation b
is short for k1/(kis), but may also be used as a design
parameter, an alternative to k1.
The PI controller may then be regarded as a PID controller with f = b = 1 and may be written as
K PI s
When it comes to implementation it is straightforward to translate these new design parameters to the traditional ones in (7).
ki
1 ss
k1 ki
s
s
p
p
4 8 2a2 a for k 1 4a 1 2 2a2 a=Gm K
95
K
K1 0.5sLd
esLd
1 sT
1 sT 1 0.5Ld
where a = T/Ld. Extensive empirical investigations, presented in e.g. [21] show that there is most often a corresponding minimum in the Jv/Ju-graph, when the
controller is a PI one. This fact is also observed in [35]
and illustrated by Fig. 5 (see also [3]).
In the optimal PID case there is normally no such
minimum. On the contrary, the Jv-graph tends to approach a horizontal asymptote as Ju grows, cf. Fig. 5.
For a non-minimum phase system it can even be proved
theoretically that a lower limit exists, see [2]. Thus it can
be argued that it is useless to increase Ju above a certain
level J uec , because the reward in terms of decreased Jv is
too small or none. This tendency towards non-decreasing Jv is more obvious and the economic level J uec is
lower the more complex the plant is (higher j value),
see [3]. The economic limit is obviously not very
sharp, but for the plant in Fig. 5 it has been estimated
to approximately 10.
It should be emphasized that an optimal PID controller, working at its economic level, mostly oers better
system properties than an optimal PI controller, and still
J uec is very reasonable. Moreover, it is seen in Fig. 5 that
when a PID controller with a = 4 [2628] and b = 10
[2931] oers the same performance Jv as the optimal
PI controller, the demanded control activity Ju is almost
5 times higher. Since these values of a and b are often
recommended and used in practice, this is one obvious
reason why derivative action is not used in most industrial applications. In some (especially older) process control systems, the parameter b is xed to b = 10 and can
not even be adjusted.
It is also worth noting that for those Ju values, where
the PI controller works at its best, the PID graphs typically come close to the PI graphs. This is what could
be expected, since then b ! 1, and a PID controller with
b = 1, f = 1 corresponds to a PI controller.
20
v
k PID
10
PID optimal
PID:a=4,b=10
PID:a=4,b opt
|Su|PID
PI optimal
Ju
2
|S |
u PI
PI
1
0.5 2
10
10
0.3s
10
Fig. 4. Plant G = e
/(1 + s) optimally controlled by a PI controller
(dashdotted), a PID controller (solid). The gure illustrates the
relations between k1 and Ju.
12
96
2.5
GM S
Among industrial users there is a widely spread scepticism against derivative controller action, based on anxiety about gain of sensor noise. This has caused the
derivative function in many controllers to be switched
o. When derivative action, despite this, is used, a
low-pass lter is always included in the control loop to
bound the HF gain. However, tuning of the controller,
e.g. for demanded stability margins, is mostly derived
without consideration of the lter, which is added afterwards. The lter factor is then typically given an ad hoc
value, low enough not to inuence the tuning. Fig. 7 offers a hint about why b =pTd/Tf P 10 has often been
chosen. Note that b b= a, why a = 4 and b = 10
approximately corresponds to b = 20. For Fig. 7 the
parameters s, f and ki have been kept constant, while
b has been changed and GMS observed for G(s) =
e0.3s/(1 + s)2. It is obvious that b has a signicant
impact on the stability margin, but that this impact
decreases for b > 20. When the lter factor is included
in the tuning parameter vector q, the properties of a system can be remarkably improved (low Jv and low Ju)
1.5
2
10
20
100
500
1/( 150G)
1.5
0.5
0.5
0
0
a
0.5
150
0
0
1
b
0.5
150
Fig. 6. Zero parameters, damping f and normalized natural frequency 1/s, as functions of j150, optimal outcomes (+) and linear approximations.
cl
97
1.5
0/150
0.5
0.5
1:st pole pair
0
0
0.2
0.4
150
0.6
0
0
0.2
0.4
150
0.6
Fig. 8. Dominating closed loop pole pairs as functions of j150 for optimal PID controllers with GMS = 1.7 and J u J uec .
Sometimes it has been claimed that when a phase margin of 45 (approximately corresponding to MT = 1.3) is
demanded, a well-tuned PID controller conserves the
phase shift of a minimum phase plant at a frequency
where it is around 135 [37,5]. This means that the controller at the open loop gain crossover frequency xc will
have no signicant phase lag. According to [37] it is even
likely to have a positive phase at this frequency.
Fig. 10 shows the phase shift Du found for optimal
PID controllers at xc for dierent values of j. It indi-
1.5
0
0
30
60
0
0.5
b/180
0.5
PID
c/180
98
1
x150G 0.44 0.86j150
s
1
2
min 3
k1
; 25
G0
j150G
x150G
0.45
ki
0.1
G0 j150 0.07
f 0.75
11
Results in terms of Jv, Ju and GMS for three dierent plants, from eight common tuning methods are
compared with each other and with the optimal controller (according to (6) with J u J uec ), see Tables 13. For
all controllers the gain is adjusted to achieve GMS = 1.7,
original values given in parentheses. When no special lter is proposed, a rst-order low-pass lter with b = 10 is
added to make the controller proper.
ZieglerNichols (ZN): ZieglerNichols method
from 1942 [26] is included because it is classical and
well-known, also among practicians and students.
H)
AstromHagglund (AH): The method named (A
is introduced in [9]. It is an empirical frequency response
method, based on dominant pole design [38]. Two demands on MS (1.4 and 2.0) are proposed. Here the mean
values of the parameters given by the two demands are
used.
Internal model control (IMC): The IMC method is
introduced in [39,40]. Since a second-order plant model
is required for the method to result in a PID controller,
a model reduction according to [41] has preceded the
parameter calculation when necessary.
The IMC controller cancels the stable plant poles by
the controller zeros, which determines f and s. The
closed loop poles are placed as a double pole (IMC)
or alternatively (IMCZ), for plants with a non-minimum
phase zero at s = z, one pole in the stable image at
s = z. The characteristic polynomials of the closed
loop poles are then P(s) = (s + 1/k)2(IMC) or P(s) =
(s + 1/k)(s + z) (IMCZ), where k is the remaining tuning
parameter.
Skogestad IMC (SIMC): This method is a
modication of IMC introduced in [8]. The included
model reduction guarantees real poles. Since b is theoretically innite, only one nite pole remains to be
placed by the characteristic polynomial P(s) = (s + 1/
k). For fast response k is recommended to equal the delay time of the reduced model, and then there is no
parameter left for adjustment. In [8] b 100 is used in
the analysis, which results in large Ju values. In the nal
implementation a smaller value is recommended.
Hence, for SIMC2 the lter factor b is adjusted to give
J u J uec .
Traditional parameter choice (ab): In Section 3 attention was called to the fact that a = Ti/Td and b = Td/Tf
in the literature very often are given the xed values 4
and 10, respectively. For this study these values are used
and Ju is bounded by J uec .
KristianssonLennartson (KLj150): This tuning
method is presented above in (11). It is based on j150
and x150G .
KristianssonLennartson (KLT63): This extremely
simple method is based on s = T63/3, f = 0.8, b = 10,
and ki tuned to fulll GMS = 1.7.
99
Table 1
Comparison of dierent methods for design of a PID controllers (the plant model is G = 1/(1 + s)4)
Method
GMS
Ju
Jv
Optimal
ZN
H
A
IMCk=1.1
SIMC
SIMC2
ab
KLj150
KLT63
1.70
1.70
1.70
1.70
1.70
1.70
1.70
1.70
1.70
8.06
15.64 (24.1)
17.8 (17.8)
2.99
74.3 (49.6)
8.06
8.06
8.06
9.07
1.62
2.01
1.82
2.16
2.00
2.16
2.76
1.64
1.71
(2.32)
(1.70)
(1.44)
(1.75)
IAE
(1.30)
(1.81)
(3.00)
(1.57)
1.93
2.01
1.83
2.32
2.35
2.55
3.49
1.90
2.10
(1.40)
(1.82)
(3.13)
(1.82)
7.05
20.0
20.4
3.46
122
14.3
21.0
7.57 (7.25)
10.0
0.70
1.03
0.95
0.80
1.03
1.03
0.98
0.75
0.80
1.65
1.57
1.59
1.87
1.22
1.22
0.97
1.67
1.45
Table 2
Comparison of dierent methods for design of a PID controllers (the plant model is G = 1/((1 + s)(1 + 0.7s)(1.0.72s)(1 + 0.73s)))
Method
GMS
Ju
Jv
IAE
Optimal
ZN
H
A
IMCk=0.66
SIMC
SIMC2
ab
KLj150
KLT63
1.70
1.70
1.70
1.70
1.70
1.70
1.70
1.70
1.70
8.67
17.7 (28.0)
20.9 (21.5)
3.40
108.4 (85.0)
8. 67
8.67
8.67
11.1
0.88
1.04
0.95
1.33
0.92
1.11
1.63
0.89
0.88
1.05
1.05
0.96
1.35
0.92
1.11
2.06
1.03
1.07
6.75
20.0
20.7
3.69
102.9
9.89
21.0
7.04 (6.68)
10.0
0.71
1.03
0.96
0.87
1.00
1.00
0.98
0.75
0.80
1.00
0.92
0.96
1.22
0.97
0.97
0.60
1.05
0.91
13.1
20.0
24.3
4.29
3.99
100
15.2
21.0
13.2 (11.5)
10.0
0.89
1.03
1.12
1.00
1.00
1.00
1.00
0.98
0.75
0.80
0.88
0.91
0.73
1.00
1.00
1.00
1.00
0.67
0.85
0.97
(2.32)
(1.73)
(1.52)
(1.77)
(0.66)
(0.93)
(1.18)
(0.86)
(0.72)
(0.93)
(1.18)
(0.98)
Table 3
Comparison of dierent methods for design of a PID controllers (the plant model is G = (1s)/(1 + s)2)
Method
GMS
Ju
Jv
Optimal
ZN
H
A
IMCk=0.61
IMCZk=0.81
SIMC
SIMC2
ab
KLj150
KLT63
1.70
1.70
1.70
1.70
1.70
1.70
1.70
1.70
1.70
1.70
6.11
8.21 (12.0)
7.64 (7.23)
1.65
1.52
41.0 (50.0)
6.11
6.11
6.11
4.52
1.98
2.21
2.32
2.61
2.62
2.44
2.48
2.49
2.26
2.23
(2.39)
(1.65)
(2.01)
(1.86)
IAE
(1.51)
(2.45)
(2.00)
(2.12)
2.59
2.63
2.80
3.03
3.04
2.84
2.90
3.20
2.98
2.87
(1.98)
(2.91)
(2.42)
(2.81)
100
J u 10 J uec . The step responses show in all cases signicantly smaller integrated errors for the PID controllers than for the PI controllers. The plants with delay all
through show better trade-o between Jv and Ju and
more favorable step responses than the corresponding
lag plants, despite the same j, GMS and Ju. However,
the prot oered by the derivative action is almost the
same for the two kinds of plant dynamics.
The reason, why derivative action is argued to be
more protable for plants with lag than for time delayed
plants, is possibly that an arbitrary linear controller,
with very large control activity Ju, theoretically can reduce Jv to zero for plants with lag. This is not the case
for time delayed plants, due to their non-minimum
phase behavior, see [7]. However, with reasonably large
control activity, the dierences between PI and PID control are comparable for plants with equal j value, independently of whether the plant dynamics include a
signicant time delay or not. Very large time delays,
on the other side, decrease the prot of derivative action, as is stated below.
Glag s
12
sLd
Gdelay s
e
1 s1 0.2s
13
1.2
PIlag
PI
0.6
=1
PID
0.4
Jv 0.5
PIdelay
PIDlag
PI
0.5
0.2
PID0.5
PIDdelay
0
0.2
10
Ju
15
20
10
3.5
20
b
0.6
PIlag
PI
L =0.4
d
0.4
PID
lag
Jv
PID
0.4
y
PI
0.2
0.2
PID
0.2
PIdelay
PIDdelay
0.5
0
c
10
Ju
15
20
0
d
6
t
Fig. 11. Left: Jv/Ju relations for Glag and Gdelay with (a) j = 0.147 (a = 0.5,Ld = 0.211) and (c) j = 0.249 (a = 1.0,Ld = 0.405). Right: Process
disturbance step responses for corresponding plants (b) Glag + PID (Ju = 10) and Glag + PI (Ju optimal), (d) Gdelay + PID (Ju = 10) and Gdelay + PI
(Ju optimal).
e
/(1 + s)
e5s/(1 + s)3
1/(1 + s)4
1/(1 + s)8
s2
Ld
J vPI (opt)
J vPID J uec
J vPID =J vPI
1
1.5
1.5
1.5
1.5
5.5
1.5
5.5
3.7
10.6
3.4
9.2
2.3
7.5
1.6
5.8
0.64
0.71
0.47
0.64
with long delay times, but also lag, there are dierent
opinions. According to [42] derivative action should be
used in those cases, while [5] means that derivative action should always be avoided for plants with signicant
time delays. In [8] a more precise recommendation says
that derivative action should only be used, when the second largest process time constant is greater than the delay time after model reduction (according to a given
scheme).
Table 4 shows a comparison of performance for optimal PI and PID controllers for some plants, where
derivative action according to [8] is not recommended
(s2 is the second largest process time constant according
to [8]). It is obvious that for all plants in the table derivative action has a contribution to oer. More general
however, there is no doubt that the improvement in performance oered by derivative action decreases when
Ld/Teq increases (Teq = eective time constant). When
Ld/Teq = 10 the improvement is approximately 15%
and for Ld/Teq = 20 it is 2% This is quite in accordance
with related results presented in [43,44].
7. Conclusions
In this paper a series of observations about closed
loop systems with optimal PI and PID controllers is presented. The expression optimal refers to controllers
that have the best possible trade-o between output performance and control activity when demanded stability
margin is guaranteed. The applied evaluation method
is briey presented in the beginning of the paper.
It has been noticed that the closed loop bandwidth xb
is very close to the phase crossover frequency of the
plant x180G and signicantly higher than the loop gain
crossover frequency xc. Furthermore, xc decreases with
j, while xb is independent. The pole placement normalized by the plant time scale is almost the same for most
optimal systems.
Observations that motivate alternative tuning parameters for the PID controller are reported. Based on
these observations simple, but still almost optimal,
tuning rules are presented. Since a PID controller
undoubtedly can oer signicantly improved performance compared to a PI controller, and the introduced
tuning rules are extremely simple to use, although they
include tuning of the low-pass lter in the derivative
101
Acknowledgments
Valuable discussions and comments on the manu strom and professor
script by professor Karl-Johan A
Sigurd Skogestad are very much appreciated.
References
[1] B. Kristiansson, B. Lennartson, Optimal PID controllers
including rollo and Smith predictor structure, in: 14th World
Congress of IFAC, vol. F, Beijing, China, July 1999, pp. 297302.
[2] B. Kristiansson, PID controllers, design and evaluation, PhD
thesis, Control and Automation Laboratory, Department of
Signals and Systems, Chalmers University of Technology, Goteborg, Sweden, August 2003.
[3] B. Kristiansson, B. Lennartson, Robust and optimal tuning of PI
and PID controllers, IEE Proc. Control Theory Appl. 149 (1)
(2002) 1725.
[4] F.G. Shinskey, Process Control Systems, Application, Design and
Tuning, McGraw-Hill, 1996.
strom, T. Hagglund, The future of PID control, in: PID00,
[5] K.J. A
IFAC Workshop of Digital Control, Past, Present and Future of
PID Control, Terrassa, Spain, April 2000, pp. 1930.
[6] B. Kristiansson, B. Lennartson, Simple and robust tuning of PI
and PID controllers, IEEE Control Syst. Mag., submitted for
publication.
[7] B. Kristiansson, B. Lennartson, Convenient, almost optimal and
robust tuning of PI and PID controllers, in: Proc. of the 15th
World Congress of IFAC, Barcelona, Spain, July 2002.
[8] S. Skogestad, Simple analytic rules for model reduction and PID
controller tuning, J. Process Control 13 (2003) 291309.
strom, T. Hagglund, PID Controllers: Theory, Design and
[9] K.J. A
Tuning, Instrument Society of America, 1995.
[10] W.K. Ho, K.W. Lim, C.C. Hang, L.Y. Ni, Getting more phase
margin and performance out of PID controllers, Automatica 35
(1999) 15791585.
[11] Q.G. Wang, C.C. Hang, Y. Yang, J.B. He, Quantitative robust
stability analysis and PID controller design, IEE Proc. Control
Theory Appl. 149 (1) (2002) 37.
[12] W.K. Ho, W. Xu, PID tuning for unstable processes based on
gain- and phase-margin specications, IEE Proc. Control Theory
Appl. 145 (5) (1998) 392396.
[13] J. Crowe, M.A. Johnson, Towards autonomous PI control
satisfying classical robustness specications, IEE Proc. Control
Theory Appl. 149 (1) (2002) 2631.
[14] R. Gorez, P. Klan, Nonmodel-based explicit design relations for
PID controllers, in: PID00, IFAC Workshop of Digital Control,
Past, Present and Future of PID Control, Terrassa, Spain, April
2000, pp. 141148.
102