Professional Documents
Culture Documents
www.elsevier.com/locate/automatica
Technical Communique
ISE tuning rule revisited
Jietae Leea , Thomas F. Edgarb;∗
a Department of Chemical Engineering, Kyungpook National University, Taegu 702-701, South Korea
b Department of Chemical Engineering, University of Texas, Austin, TX 78712-1062, USA
Received 28 April 2003; received in revised form 19 January 2004; accepted 10 March 2004
Abstract
The integral square error (ISE) method is applied to approximate the internal model control (IMC) system by the PID control system.
Simulations show that the proposed method to design PID controllers can relieve some disadvantages of the original ISE method and
IMC-based PID controller design methods.
? 2004 Elsevier Ltd. All rights reserved.
Keywords: Proportional-integral-derivative control; Integral square error; Internal model control; Frequency domain design
1. Introduction control system instead of the set point. A new cost functional
to be minimized is de8ned and the disadvantage of oscilla-
By applying single-input single-output (SISO) tun- tory responses of the ISE method is somewhat relieved. The
ing methods sequentially, multiloop control systems for ISE tuning method is a nonlinear problem and requires an
multi-input multi-output (MIMO) processes can be de- iterative solution. To reduce the computational complexity,
signed (Mayne, 1973). This requires designing PID con- we convert the ISE tuning problem to solving a weighted
trollers for processes not represented by standard 8rst or least-squares problem.
second order plus time delay models. Optimization-based
methods such as the integral square error (ISE) method
2. ISE tuning method
can be applied to such processes without needing model
reduction. Here an ISE method with reduced computational Many formulas for the design of PID controllers have
complexity is proposed. been published (Seborg et al., 2003; Astrom & Hagglund,
Compared to the integral time absolute error (ITAE) tun- 1995). Here, optimization-based methods which can be ap-
ing method, the ISE tuning method for the step set point plied to general process models are considered. Among
change shows oscillatory responses and hence it is not pre- them, the ISE method based on the criterion
ferred generally (Seborg, Edgar, & Mellichamp, 2003). That ∞
is partly because the ISE method forces the process output J1 = (ysp (t) − y(t))2 dt (1)
to follow the abrupt set point change (Astrom & Hagglund, 0
1995). This abrupt error cannot be eliminated e<ectively has some advantages such as simple computation in the fre-
by a feedback controller. To avoid this disadvantage, the quency domain (Chen, 1978). Here y(t) and ysp (t) are the
ITAE method introduces the weight of time which puts less output and its set point, respectively. The ISE tuning rule
weight on the initial large error. Alternatively, we force the often provides oscillatory results because of initial large er-
process output to follow the output of the internal model rors that cannot be eliminated e<ectively by the feedback
controller (Astrom & Hagglund, 1995). Hence, we modify
This paper was not presented at any IFAC meeting. This paper was
the criterion slightly as follows:
recommended for publication in revised form by Associate Editor Jay H. ∞
Lee under the direction of Editor Paul Van den Hof. J2 = (yd (t) − y(t))2 dt; (2)
∗ Corresponding author. Tel.: +1-512-371-3080; fax: +1-512-
0
471-7060.
E-mail addresses: jtlee@knu.ac.kr (J. Lee), tfedgar@che.utexas.edu where yd (t) is the desired output variable which can be
(T.F. Edgar). obtained by a physically realizable feedback controller.
0005-1098/$ - see front matter ? 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.automatica.2004.03.008
1456 J. Lee, T.F. Edgar / Automatica 40 (2004) 1455 – 1458
3. IMC–ISE tuning method dynamics are not activated, PID controllers can approximate
CIMC (s) and small HC(j!) makes Eq. (4) e<ective. Dis-
For the desired output of yd (t), the output of the con- cretizing Eq. (4), we have a cost function
trol system designed by the internal model control (IMC)
H (j!)(1 − H (j!)) 2
method (Morari and Za8riou, 1989) or equivalently by the J3 = I I 2
!1:5 CIMC (j!) C IMC (j!) − C(j!) ;
direct synthesis method (Seborg et al., 2003) is used here. !∈
Applying Parseval’s theorem to Eq. (2) (5)
∞ where is a set of frequencies with equal logarithmic
1 spacing.
J2 = Yd (j!) − Y (j!)2 d!
2 −∞ I
For a PID controller, C(s) = sC(s) = k1 + kP s + kD s2 ,
minimization of Eq. (5) is a weighted linear least-squares
1 ∞ G(j!)CIMC (j!)
= problem and can be solved easily without iteration. For a
0 j!(1 + G(j!)CIMC (j!)) I
PID controller with a 8lter, C(s) = sC(s) = (kI + kP s +
2 2
kD s )=(1 + F s), J3 becomes a nonlinear least-squares prob-
G(j!)C(j!)
− d!; (3) lem but eJcient methods to solve it are available (for ex-
j!(1 + G(j!)C(j!))
ample, the routine “invfreqs” in the Signal Processing Tool-
box for MATLAB (Math Works, 1997)). As in the IMC
where method, is used as a design parameter. The frequency set
of is chosen as 101 frequencies which have equal spacing
1 H (s) logarithmically between 0:01= and 10=.
CIMC (s) = ;
G(s) 1 − H (s) This IMC–ISE method can be extended to discrete-time
control systems. Consider a discrete-time cost function
G + (s)
H (s) = : ∞
(s + 1)r J˜ 2 = (yd (kT ) − y(kT ))2 ; (6)
k=0
G + (s)=non-minimum phase part of G(s) and G + (j!)=1 where T is a sampling time. Applying the discrete-time
for any real ! Parseval’s theorem (Ogata, 1987)
r=relative order between the numerator and denominator 2=T
in G(s)=G + (s). ˜ T
J2 = Ỹ d (ej!T ) − Ỹ (ej!T )2 d!; (7)
The cost function of Eq. (3) is nonlinear for the unknown 2 0
C(s), and requires an iterative minimization method. To where Ỹ (z) and Ỹ d (z) are the z transformation of y(kT )
simplify the computations, we rewrite Eq. (3) using C(s) = and yd (kT ), respectively. A least-squares problem similar
CIMC (s) + HC(s) to Eq. (5) can be obtained by applying the above technique
2 to Eq. (7).
1 1 GCIMC
∞
G(CIMC + HC) d!
J2 = −
0 ! 1 + GCIMC 1 + G(CIMC + HC) !
2 4. Examples
1 ∞ 1 G
HC2 d!
≈
0 ! (1 + GCIMC ) 2 ! The IMC–PID method (Morari & Za8riou, 1989) and the
2 desired closed-loop response (DCLR) method (Lee, Park,
1 ∞ H (j!)(1 − H (j!)) Lee, & Brosilow, 1998) are both based on the internal
=
0 !1:5 CIMC (j!) model controller and used to design PID controllers. The
IMC–PID method requires model reduction when applied to
d!
×CI IMC (j!) − C(j!)
I 2
; (4) higher-order processes and hence is not considered here. The
! proposed method (IMC–ISE) is compared with the DCLR
method and ISE optimal method with a cost of Eq. (2). The
I
where CI IMC (s) = sCIMC (s) and C(s) = sC(s). The approx- DCLR method uses Maclaurin series expansion of CIMC (s),
imation in Eq. (4) is based on GC=(1 + GC) = G(CIMC + resulting in an accurate approximation at low frequency.
HC)=(1 + G(CIMC + HC)) = GCIMC =(1 + GCIMC ) + However, for some processes shown in examples below, the
G=(1 + GCIMC )2 HC + G 2 =(1 + GCIMC )3 HC 2 + · · · CIMC (s) DCLR method fails to provide PID parameters with positive
is PID-type for second-order processes and, for processes values.
whose dominant dynamics are of the second order, the
magnitude of HC(j!) will be small and the above approx- Example 1. First a 8rst-order plus time delay model,
imation will work well. Hence, when the IMC 8lter time G(s) = exp(−s)=(s + 1), is considered. Table 1 shows
constant is not too small so that higher order process ISE values for the unit step change. We can see that the
J. Lee, T.F. Edgar / Automatica 40 (2004) 1455 – 1458 1457
Table 1
Integral of square errors
1.5 1.5
1
1
0.5 0.5
0 0
0 5 10 0 20 40
time time
Fig. 1. Step set point responses for the Example 1 with = 0:5. (Dotted Fig. 2. Step set point responses for the Example 2. (Dotted line: desired
line: desired output, dashed line: DCLR controller, solid line: IMC–ISE output, dashed line: DCLR controller, solid line: IMC–ISE controller).
controller).
IMC–ISE method provides controllers with slightly better and the proposed method can also be applied to obtain con-
ISE values than the DCLR method and, as the time delay troller parameters for load changes.
increases, the IMC parameter becomes larger (==3) com-
pared to the process time constant, resulting low e<ective Example 2. Consider a process with dominant lead time
frequencies and smaller di<erence in ISE values between the constant (Lee et al., 1998)
DCLR and proposed methods. Fig. 1 shows step set point
s2 + 2s + 0:25
responses and load responses for the unit step disturbance G(s) = : (8)
entering at the process input. Compared to the set point re- s4 + 6:5s3 + 15s2 + 14s + 4
sponses, both load responses are somewhat poor. The IMC Due to the dominant zero, the DCLR method provides
control is based on the pole/zero cancellation and its poor PID controllers with negative gains. By introducing a 8lter
load response is well-known (Astrom & Hagglund, 1995). to cancel the zero, a positive controller gain can be obtained.
A two-degree-of-freedom controller can be used for both The IMC–ISE method yields positive parameters regardless
set point and load changes. The DCLR method (Lee et al., of 8lters. Table 1 shows ISE values for the unit step change
1998) provides di<erent tuning rules for load disturbances and Fig. 2 shows step set point responses and load responses
1458 J. Lee, T.F. Edgar / Automatica 40 (2004) 1455 – 1458
for the unit step disturbance entering at the process input for 1.5
control systems of PID with 8lter designed by the IMC–ISE
method and the DCLR method. Far better control systems
with lower ISE values can be obtained by the proposed
method. 1