Documente Academic
Documente Profesional
Documente Cultură
TRANSACTIONS®
ISA Transactions 41 共2002兲 511–520
Abstract
Practical requirements on the design of control systems, especially process control systems, are usually specified in
terms of time-domain response, such as overshoot and rise time, or frequency-domain response, such as resonance peak
and stability margin. Although numerous methods have been developed for the design of the proportional-integral-
derivative 共PID兲 controller, little work has been done in relation to the quantitative time-domain and frequency-domain
responses. In this paper, we study the following problem: Given a nominal stable process with time delay, we design
a suboptimal PID controller to achieve the required time-domain response or frequency-domain response for the
nominal system or the uncertain system. An H ⬁ PID controller is developed based on optimal control theory and the
parameters are derived analytically. Its properties are investigated and compared with that of two developed suboptimal
controllers: an H 2 PID controller and a Maclaurin PID controller. It is shown that all three controllers can provide the
quantitative time-domain and frequency-domain responses. © 2002 ISA—The Instrumentation, Systems, and Auto-
mation Society.
Keywords: Linear system; Time delay; PID controller; Optimal control; Time-domain response
0019-0578/2002/$ - see front matter © 2002 ISA—The Instrumentation, Systems, and Automation Society.
512 Weidong Zhang et al. / ISA Transactions 41 (2002) 511–520
K ⫺ ⫹K ⫹ ⫺⫹ ⫹ ⫺⫹ ⫹ 共 s⫹1 兲共 1⫹ s/2兲
K⫽ , ⫽ , ⫽ . Q im 共 s 兲 ⫽ .
2 2 2 K
A central objective in automatic control is that a Obviously Q im ( s ) is improper. Now we use the
physical quantity is made to behave in a pre- following low-pass filter to roll Q im ( s ) off at high
scribed way by using the error between the system frequency:
output and the setpoint input. This gives rise to the 1
optimal control. Assume that the optimal perfor- J共 s 兲⫽ , ⬎0.
mance index is H ⬁ optimal, i.e., min储W(s)S(s)储⬁ , 共 s⫹1 兲 2
where W ( s ) is a weighting function. W ( s ) should Then
be selected such that the 2-norm boundary of the
system input is normalized by unity. That is, 共 s⫹1 兲共 1⫹ s/2兲
W ( s ) ⫽1/s for a unit step setpoint. The perfor- Q 共 s 兲 ⫽Q im 共 s 兲 J 共 s 兲 ⫽ .
K 共 s⫹1 兲 2
mance index implies that the controller is designed 共11兲
to minimize the worst error resulting from system
inputs. With Padé approximation, we have As tends to be zero, the controller tends to be
optimal. The corresponding controller of the unity
1⫺ s/2 feedback loop is
G 共 s 兲 ⫽K , 共7兲
共 s⫹1 兲共 1⫹ s/2兲 1 共 s⫹1 兲共 1⫹ s/2兲
C共 s 兲⫽ . 共12兲
which will be regarded as the nominal plant uti- Ks 2 s⫹2⫹ /2
lized to derive the H ⬁ PID controller. The error
introduced by the approximation is included in un- The suboptimal H ⬁ PID controller is derived ana-
certainty, which will be discussed later. lytically. Comparing the above controller with the
Instead of a numerical method, an analytical de- following practical PID controller,
sign procedure is developed for the given plant. It
is seen that W ( s ) S ( s ) is analytical in the open
right half plane. According to the well-known
冉
C 共 s 兲 ⫽K C 1⫹
1
T Is
⫹T D s 冊 1
T F s⫹1
. 共13兲
maximum modulus theorem, a fundamental fact The parameters of the new PID controller are
concerning complex functions, 兩 W ( s ) S ( s ) 兩 does
not attain its maximum value at an interior point 2
T F⫽ , T I⫽ ⫹ , T D⫽ ,
of the open right half plane. On the other hand, the 2⫹ /2 2 2T I
G ( s ) has a zero at s⫽2/ in the open right half
plane. Thus, for all Q ( s ) ’s, TI
K C⫽ .
K 共 2⫹ /2兲
储 W 共 s 兲共 1⫺G 共 s 兲 Q 共 s 兲兲 储 ⬁ ⭓ 兩 W 共 2/ 兲 兩 . 共8兲
If the practical PID controller is in the form of
冉 冊
Consequently we have
1 T D s⫹1
min储 W 共 s 兲 S 共 s 兲 储 ⬁ ⫽min储 W 共 s 兲共 1⫺G 共 s 兲 Q 共 s 兲兲 储 ⬁ C 共 s 兲 ⫽K C 1⫹ , 共14兲
T I s T F s⫹1
⫽ /2. 共9兲 the parameters of the PID controller are
However, W ( s ) has a pole on the imaginary axis. 2
To obtain a finite infinity norm, a constraint will T F⫽ , T I⫽ , T D⫽ ,
2⫹ /2 2
be imposed on the design procedure:
lim 共 1⫺G 共 s 兲 Q 共 s 兲兲 ⫽0. 共10兲 TI
K C⫽ .
s→0 K 共 2⫹ /2兲
514 Weidong Zhang et al. / ISA Transactions 41 (2002) 511–520
冋 册
T D⫽ ⫺T F , K C⫽ .
2T I K 共 2⫹ /2兲 1 f ⬙共 0 兲 2
C共 s 兲⫽ f 共 0 兲 ⫹ f ⬘ 共 0 兲 s⫹ s ⫹¯ .
Since Eq. 共13兲 is used in many papers, for com- s 2!
parison, only this form is considered in the later 共18兲
part of this paper.
Taking the first three terms, we obtain the Maclau-
If a conventional PID controller is installed, the
rin PID controller, of which the parameters are 关9兴
parameter T F has been determined. Usually T F
⫽0.1T D 关4兴. In this case, one can also use the
2
above rules by simply omitting the new T F . Then, T F ⫽0, T I⫽ ⫹,
similar results will be obtained. 2 共 ⫹ 兲
冋 册
Both the H 2 PID controller and Maclaurin PID
controller are based on the result of IMC. In IMC, 2 TI
T D⫽ 1⫺ , K C⫽ .
the plant is factored as 2 共 ⫹ 兲 3T I K 共 ⫹ 兲
G 共 s 兲 ⫽G ⫹ 共 s 兲 G ⫺ 共 s 兲 , Sometimes, one may use a second-order model.
where G ⫹ ( s ) contains all nonminimum phase fac- The above methods can be directly extended to
tors, G ⫹ ( 0 ) ⫽1, and G ⫺ ( s ) is the minimum phase this case. Since the design procedure is almost the
portion of the model. The IMC controller is then same, only the results are given here. Suppose that
given by the following formula 关8兴: the model is
1 Ke ⫺ s
Q共 s 兲⫽ , G共 s 兲⫽ . 共19兲
共 s⫹1 兲 n G ⫺ 共 s 兲 共 1 s⫹1 兲共 2 s⫹1 兲
where n is chosen so that the controller is bi- Utilizing the Taylor series, the PID controller pa-
proper, that is, both Q ( s ) and 1/Q ( s ) are proper. rameters designed by the proposed method are
For the given plant we have
2 1 2
s⫹1 T F⫽ , T I⫽ 1⫹ 2 , T D⫽ ,
C共 s 兲⫽ , 共16兲 2⫹ TI
s⫹1⫺e ⫺ s
Since a time delay is included in the controller, TI
K C⫽ . 共20兲
it cannot be directly implemented. Many methods K 共 2⫹ 兲
have been presented to approximate the control-
ler by a rational transfer function, such as a nu- The H 2 PID controller parameters are
merical method, Padé approximation, Taylor se-
ries, Maclaurin series, and so on 关11兴. The H 2 sub- 1 2
optimal PID controller given by 关7兴 utilizes a Padé T F⫽ , T I⫽ ,
⫹2 1⫹ 2
approximation. The resultant controller is
1 共 s⫹1 兲共 1⫹ s/2兲 TI
C共 s 兲⫽ 共17兲 T D⫽ 1⫹ 2 , K C⫽ . 共21兲
Ks s/2⫹⫹ K 共 ⫹2 兲
and the PID controller parameters are The Maclaurin PID controller parameters are
Weidong Zhang et al. / ISA Transactions 41 (2002) 511–520 515
2 2 ⫺ 2 and
T F ⫽0, T I⫽ 1⫹ 2⫺ ,
2 共 2⫹ 兲 1⫺0.5s
T共 s 兲⫽ .
2 2 ⫺ 2 1 2 ⫺ 3 / 共 12⫹6 兲 共 s⫹1 兲共 1⫹0.5s 兲
T D ⫽⫺ ⫹ ,
2 共 2⫹ 兲 TI Only one pole of the nominal plant is canceled.
Note that there is an adjustable parameter in
TI the three PID controllers. It has been shown by
K C⫽ . 共22兲
K 共 2⫹ 兲 some researchers that relates directly to the
nominal performance and robustness of the sys-
3. Discussion tem. This paper will show that the quantitative
time-domain response and frequency-domain re-
All three PID controllers are suboptimal. The sponse can be gained by adjusting .
design procedure shows that they relate closely Consider the H ⬁ PID controller again. We re-
and each has its own features. The difference be- gard the error introduced by the approximation as
tween the Maclaurin PID controller and the H 2 uncertainty. The actual plant is in the form of a
and H ⬁ PID controllers is obvious. The difference first-order plus time delay. Then
between the H 2 PID controller and the H ⬁ PID
共 1⫹ s/2兲 e ⫺ s
controller lies in the manner they cancel the poles T共 s 兲⫽
of the plant. The proposed H ⬁ controller tends to 2 s 2 ⫹ 共 2⫹ /2兲 s⫹ 共 1⫹ s/2兲 e ⫺ s
cancel all poles of the process while the H 2 con- 共23兲
troller tends to cancel the poles of the minimum and
phase part of the process.
2 s 2 ⫹ 共 2⫹ /2兲 s
S共 s 兲⫽ .
3.1. Example 1 2 s 2 ⫹ 共 2⫹ /2兲 s⫹ 共 1⫹ s/2兲 e ⫺ s
共24兲
The difference between the H 2 PID controller
and the H ⬁ PID controller is illustrated in this ex- It is seen that T ( s ) does not depend on K and .
ample. Consider the plant described by This implies that the setpoint response for the
closed-loop system relates only to and . The
e ⫺s similar result also exists for the H 2 PID controller.
G共 s 兲⫽ . Strictly speaking, the setpoint response of the sys-
s⫹1
tem with the Maclaurin PID controller relates to
The nominal plant with Padé approximation is not only and but . Since the difference is very
1⫺0.5s
G共 s 兲⫽ .
共 s⫹1 兲共 1⫹0.5s 兲
The H ⬁ method yields that
共 s⫹1 兲共 1⫹0.5s 兲
Q共 s 兲⫽
共 s⫹1 兲 2
and the nominal complementary sensitivity trans-
fer function is
1⫺0.5s
T共 s 兲⫽ .
共 s⫹1 兲 2
We see that both of the two poles of the nominal
plant are canceled. For the H 2 controller we have
s⫹1 Fig. 3. Responses of system A. 共Solid line: H ⬁ ; dashed
Q共 s 兲⫽
s⫹1 line: Maclaurin; dotted line: H 2 .兲
516 Weidong Zhang et al. / ISA Transactions 41 (2002) 511–520
Fig. 4. Responses of system B. 共Solid line: H ⬁ ; dashed Fig. 6. Quantitative rise time. 共Solid line: H ⬁ ; dashed line:
line: Maclaurin; dotted line: H 2 .兲 Maclaurin; dotted line: H 2 .兲
Fig. 5. Quantitative overshoot. 共Solid line: H ⬁ ; dashed Fig. 7. Quantitative ISE. 共Solid line: H ⬁ ; dashed line:
line: Maclaurin; dotted line: H 2 .兲 Maclaurin; dotted line: H 2 .兲
Weidong Zhang et al. / ISA Transactions 41 (2002) 511–520 517
above analysis. We define as the ‘‘performance that the ratio of ISE and is also determined by
degree.’’ It is seen that the overshoot and the ratio the performance degree 共Fig. 7兲. This implies that
of rise time and are determined only by the ratio the relationship between the classical performance
of the performance degree and 共Figs. 5 and 6兲. indices and the optimal performance indices is es-
The break in Fig. 6 is caused by the difference of tablished.
the definition of the rise time for systems with In frequency-domain analysis, one important
overshoot and without overshoot. The empirical concept is the resonance peak, that is, the maxi-
formulas for estimating the two indices can also be mum modulus of the closed-loop transfer function.
given. For example, the formulas of the H ⬁ PID In 关14兴, it is referred to as the maximum log
controller are as follows: modulus and utilized to design PID controller. A
commonly used specification for it is 2dB. The
Overshoot
再
quantitative relationship between the resonance
⫺0.86共 / 兲 3 ⫹14.21共 / 兲 2 peak and the performance degree is shown in Fig.
8.
⫽ ⫺8.72/ ⫹1.86, 0.1⭐/ ⭐0.59 Another important concept in frequency domain
0, 0.59⭐/ ⭐1.2, is stability margin, on which numerous methods
are based 共see, for example, 关15兴兲. The magnitude
共25兲 stability margin and phase stability margin provide
Rise time intuitive tools for control system design and are
再
very familiar to engineers. The quantitative rela-
30.34共 / 兲 3 ⫺24.63共 / 兲 2 tionship between the stability margin and the per-
formance degree is shown in Figs. 9 and 10.
⫽ ⫹8.48/ ⫺0.45, 0.1⭐/ ⭐0.59
Recently, 关16兴 proposed a new frequency-
3.97/ ⫺1.02, 0.59⭐/ ⭐1.2. domain index for PID controller design,
共26兲 1/max兩 Re关 G 共 j 兲 C 共 j 兲兴 兩 ,
In 关12,13兴, nonovershoot and monotone nonde- which involves both the magnitude stability mar-
creasing response is studied. For the three control- gin and phase stability margin to a certain extent.
lers it is very easy to get such a response. It is found that there also exists the quantitative
Both the H 2 control and the H ⬁ control relate to relationship between the new index and the perfor-
the integral square error 共ISE兲. The H 2 control mance degree 共Fig. 11兲.
minimizes the ISE for a particular input and the Unfortunately, the simple quantitative relation-
H ⬁ control minimizes the worst ISE resulting ship comes into existence only for setpoint re-
from any two-norm bounded inputs. It is found sponse. The transfer function from the load distur-
Fig. 8. Quantitative resonance peak. 共Solid line: H ⬁ ; Fig. 9. Quantitative magnitude margin. 共Solid line: H ⬁ ;
dashed line: Maclaurin; dotted line: H 2 .兲 dashed line: Maclaurin; dotted line: H 2 .兲
518 Weidong Zhang et al. / ISA Transactions 41 (2002) 511–520
冏 冏
The quantitative uncertainty profile that guaran-
G̃ 共 s 兲 ⫺G 共 s 兲 tees robust performance for an H ⬁ PID controller
⫽ 兩 ⌬ m 共 j 兲 兩 ⭐⌬ 共 兲 , 共27兲 is shown in Fig. 13.
G共 s 兲
In classical control theory, the controller is usu-
ally designed for the nominal performance speci-
fication and used for the control of the uncertain
plant. If a good model can be obtained, the esti-
mates given by the above study are in good agree- the H ⬁ PID controller, the relationship is exact,
ment with the actual time-domain responses. If the and for the Maclaurin PID controller, the relation-
uncertainty scope is large, one may hope to know ship roughly exists. Thus, all three controllers can
whether the required performance is obtained or be designed for desired time-domain response or
not. Eqs. 共28兲 and 共29兲 give a perfect estimate. frequency-domain response. The given quantita-
However, they are more mathematical than practi- tive relationship between the classical perfor-
cal and very inconvenient. As a matter of fact, we mance indices and the optimal performance indi-
usually wish to estimate the ‘‘worst case’’ re- ces makes it possible to build a bridge between
sponse 共i.e., the gain and time delay are at their the classical design method and modern design
upper limits, while the time constant is at its lower methods.
limit兲 or the range of the closed-loop response. For practical purposes, the H 2 PID controller
关17兴 recently developed a method for designing may be the most convenient one, because it has a
IMC systems with parametric uncertainty. First, a linear relationship with many performance indices.
specified value M p is chosen based on the maxi- Since there always exists uncertainty in practice,
mum desired overshoot, or the ‘‘worst case’’ over- the quantitative design for the system with para-
shoot. Second, the magnitude of the complemen- metric uncertainty is also discussed, and a new
tary sensitivity transfer function is designed to be design method is proposed to estimate the ‘‘worst
equal to or less than M p. For example, for the case’’ response or the range of the closed-loop re-
‘‘worst case’’ overshoot 10%, M p⫽1.05. By solv- sponse of the proposed controller.
ing 兩 T ( j ) 兩 ⭐M p, a unique solution on can be It is also shown that the suboptimal PID gives a
obtained. satisfactory approximation to the exact time delay
Here, an alternative method is proposed to esti- compensated scheme. This implies that the PID
mate the ‘‘worst case’’ response or the range of the controller can provide relatively good response for
closed-loop response of the proposed controller. systems with time delay, even when the time delay
Assume that we hope to achieve the ‘‘worst case’’ is very large.
overshoot of 5%. The new procedure is as follows:
共i兲 Design the controller for the nominal Acknowledgments
plant. For 5% overshoot, ⫽0.5.
共ii兲 Substitute the nominal plant by the This project was supported by the National
⫹
‘‘worst case’’ plant K ⫹ e ⫺ s / ( ⫺ s⫹1 ) . Natural Science Foundation of China 共Grant
共iii兲 Increase the performance degree No. 69804007兲 and the National Key Technologies
monotonically until the overshoot is R&D Program in the Tenth Five-Year Plan 共Grant
equal to 5%. No. 2001BA201A04兲.