SICE-ICASE International Joint Conference 2006

Oct. 18-2 1, 2006 in Bexco, Busan, Korea

Design of Robust PID Controllers for Unstable Processes

M. Shamsuzzohal, Moonyong Lee2*
School of Chemical Engineering and Technology, Yeungnam University, Kyongsan, 712-749, Korea
l(Tel. :+82-53-810-3241; E-mail: smzoha2002ghotmail.com)
*2(Tel. :+82-53-810-2512; E-mail: mynleegyu.ac.kr)
Abstract: The IMC-PID tuning rule has been proposed for the several unstable processes and unstable process with
negative zero. The results of the searching of the IMC filter for the PID controller design clearly exhibit that the
critically damped filter doesn't always gives the best response, and it depends upon the category of the process model.
The value of damping factor (i) in IMC filter can be adjusted as the process having different time-delay/time-constant
ratio for getting optimum value of integral time. The controller setting gives a robust performance for uncertainty in the
process model parameters. The simulation studies show that the proposed design method provides better disturbance
rejection when the controller are tuned to have the same degree of robustness by a measure of maximum sensitivity
(Ms). The set-point response has been adjusted by using the setpoint filter of as a two degree of freedom controller.
Keywords: IMC-PID; Robust control; Unstable process
simulation examples, and Yang et al. [8] and Liu et al.
1. INTRODUCTION [10] pointed out that executable high order controllers
The proportional integral derivative (PID) controller could result in much better system performance and
algorithm is undoubtedly the most adopted controllers robustness in comparison with conventional PID
for industrial plants, mainly due to their simplicity, and controllers. In addition, 2DOF control methods based on
they can assure satisfactory performances for a wide the Smith-predictor (SP) had been proposed by Majhi &
range of processes. Moreover, as reported [1-10] Atherton [5]; Zhange et al. [9] and really achieved
majority of the controllers used in the process control smooth nominal setpoint response without overshoot for
applications are of the PID type. It is difficult to get the first order unstable processes with time delay. It is a
similar cost/benefit ratio from other controllers to PID. notable merit that the nominal setpoint response tends to
It is well-known that control system design for an open- be faster without overshoot for unstable processes
loop unstable process is more difficult than that for a according to either the modified IMC methods or the
stable one because of the unstable nature of the modified SP methods. In fact, the common
dynamics. The some of the examples are the characteristic of the abovementioned modified IMC and
polymerization furnaces, continuous stirred tank SP methods is utilizing the nominal process model in
reactors (CSTRs), and the batch chemical reactor, which their control structures, which effectively contributes to
has a strong nonlinearity due to heat generation term in acquire the above merit. It should be noted that most
the energy balance. Recently, tuning of controllers for a existing 2DOF control methods restricted attention on
time-delay unstable process has been an active area of unstable processes modeled in the form of a first order
research in the literature. rational part plus time delay, which in fact, cannot
Huang and Chen [1] suggested a three-element structure, represent a variety of industrial and chemical unstable
which is equivalent to a two-degrees-of-freedom (2DOF) processes well enough. Besides, there usually exist the
control scheme, for controlling the open-loop unstable process unmodeled dynamics that inevitably tend to
processes. However, PID controller tuned by their deteriorate the control system performance severely for
method still gave about 100% overshoots to a setpoint unstable processes, especially for the load disturbance
change. rejection. Many existing methods, however, failed to
Lee et al. [2] and Wang and Cai [3] had proposed two- provide the on-line tuning rule of the controller
degree-of-freedom control methods for several parameters to cope with the process uncertainty.
processes to overcome excessive overshoot and large Therefore, in the proposed study the primary goal to
settling time in setpoint response and really achieved keep the control system structure simple as a
enhanced system performance. conventional feedback, which gives the best disturbance
In view of that the IMC structure is very powerful for rejection. However the setpoint response is usually
controlling stable processes with time delay and cannot accompanied with excessive overshoot, that effect can
be directly used for unstable processes by reason of the be eliminated by using the set-point filter.
internal instability Morari and Zafiriou [6], some The main objectives of the proposed study are:
modified IMC methods of 2DOF control such as Huang (i) To design the IMC-PID controller for the several
& Chen [1]; Tan et al. [7], Liu et al. [10] had been unstable processes. (ii) Robust stability analysis based
developed for controlling unstable processes with time on norm bound principle. (iii) Closed-loop time constant
delay and therefore, noteworthy improvement had been (i) guidelines (iv) Comparison of the proposed tuning
obtained for both of the setpoint tracking and the load rules with other tuning methods.
disturbance rejection, among which Liu et al. [10]
method had shown the best superiority by using the

89-950038-5-5 98560/06/$10 C 2006 ICASE

3324
 2. THEORY The filter for the IMC controller can be designed to
Figure l(a) shows the block diagram of IMC control satisfy two criteria. One is to make the IMC controller
proper and the other is to cancel both the stable and or
structure, where GP is the process model and q is the unstable poles near the zero of GD .
IMC controller. The controlled variable are related as
(7)
C= Gpq
I + q (Gp -
R+ -
Gpq GDd (1)
r
ES2-+ 2As + 1)n
~p ) I + q (Gp - Gp ) Disturbance d

For the nominal case (i.e., GP = GP ), the setpoint and IMO GD

controller Process
disturbance responses are simplified as:
C (2) setpoint
C = Gpq
RGq
C= [- Gpq]GD (3)
d Fig. 1 (a) IMC Structure
The classical feedback control structure as shown in Fig.
1 (b), the setpoint and disturbance responses are
represented by
C G6Gp (4) setpoint

R 1+GCGP
C= GD (5)
Fig. 1 (b) Classical Feedback Control
d I+G GG
where Gc denotes the feedback controller. where fi is obtained to cancel the unstable poles of
The IMC (Fig. 1 (a)) has been shown to be a powerful GD and m is the number which can be adjusted to
method for control system synthesis [6]. However, for make the IMC controller proper. Equation (7) is
unstable processes the IMC structure cannot be function as a filter with adjustable time constant A and
implemented exactly similar to stable process, since any
damping coefficient 4.
input d (s) will make C/R grow without bound if Gp is
-Gq s=dupi, du.. =0 (8)
unstable. Nevertheless, as discussed in [6], we could
still use IMC approach to design a controller for an where dup, # 0.
unstable process, if only the following conditions are
satisfied for the internal stability of the closed-loop Thus, the IMC controller is
system: -1 li= + (9)
(i) q stable (A2s2 + 2AJs + 1)
(ii) Gpq stable
Then, we get
(iii) (1- Gpq) G stable
(10)
These conditions result in the well known standard C
iPlA+1)
Gq (=
interpolation conditions [6]. If above said condition are R (22S2 + 22As + 1)m
satisfied, then the closed-loop response for setpoint as
well as load change becomes stable. (11)
2.1 IMC controller design step C=(1 G,q)GD = I1-PA (2 2 ) GD
The IMC controller design involves two steps:
The lead term (m 8is +I) in Eq. (7) causes an
Step 1: A process model GP is factored into invertible
and non invertible parts overshoot in the closed-loop response to a setpoint
change. This problem can be resolved if we add a
GP = PMPA (6) setpoint filter.
where P. is the portion of the model inverted by the I (12)
controller; PA is the portion of the model not inverted
fR

(Y1fiS +1)

by the controller (it is usually a non-minimum phase The resulting IMC controller in Eq. (9) has a stable
and contains dead times and/or right half plane response and a classical feedback controller equivalent
zeros);PA (0) =1. to IMC can be obtained from the following relationship
Step 2: The IMC controller is set as q = PM 1f . Here, G, = (13)
c 1-G q
q has zeros at up1, ..., upk because PM-1 is the inverse Thus the feedback controller is given as:
of the model portion with unstable poles.

3325
 maximum and the maximum peak for is 1. Therefore,
PA 1 (E- i= 1
(14)
G =-
(2S2 + 22;s +1) in practice choosing filters with 4 . 0.707 have no
inherent problem of the stability for designing the IMC
I _ pM pApM -1 (i=+ ) controller. The damping factor J = 0.50 minimizes
(i2s2 + 2AJs + 1)m the ISE (Morari & Zafiriou [6]). However, with this
The resulting closed-loop output response in Eq. (14) is
physically realizable, but it does not have the standard filter If =1.15; thus, performance improvement occurs
PID controller form. Therefore the next step is to design at the expense of a reduced robustness margin. So,
the PID controller that most closely approximates the choosing filters with < 0.707 in Eq. (21) is usually not
equivalent feedback controller. Lee et al. [4] proposed worthwhile.
an efficient method for converting the ideal feedback 2.3 Proposed tuning rule
controller G to a standard PID controller. Since G, has First Order Delayed Unstable process (FODUP)
an integral term, it can be expressed Consider a first order delayed unstable process of the
G f(s) (15) form:
s

Expanding GC in Maclaurin series in s gives GGP =G

= D =(22)Ke6o
rs -1
(2
(16) Where K is the gain, T is the time constant and 0 is
G, = IKf (°)+f (O)s+ f)s2+ ...

the time delay. The proposed general type of IMC filter

The first three terms of the above expansion can be is found as f = (g8s + 1)/(,2S2
2s2s
+ + 1) . Then, the
interpreted as the standard PID controller given by
resulting IMC controller becomes
( J+rs (17)
G =KK1+1
q=(is-1)(,Ps+1)1K(2L72 +2As+1) . Therefore, the ideal
where feedback controller equivalent to the IMC controller is
K, = f(O) (I18a) (/(s-1)(,8Os+
1) (23)
G, =

zC = f (0)/f(0) (I18b) K[(22s2 +22s+1) -e-s (,8is+1)]

TD= f (0o)/2f (0o) (1 8c) Expanding dead time e-os in Maclaurin series
In this way, using Eqs. (18a)-(1 8c) we can convert the e- =1-8s+#s2/2-63s3/6 and substituting in Eq. (23)
ideal controller given by Eq. (14) to a standard PID and after simplification, we have
controller. (24)
2.2 Filter selection -1+( -,)s+45s+...
In order to get the zero offset to step inputs, we adopt K[(0-,+24)s+(2 6+I2)s +(0 X SX ]
the following convention for PA and filter
The analytical PID tuning formula can be given from Eq.
f (s) = p (s)/q (s) (18) as
PA = P (°) = q (O) = I (19) Kc K= (-+2 ) (25a)
The simplest filter f (s) satisfying Eq. (7) is of the
form =~(r+3)- (2 _o2 +1* ) (25b)

f= (1)/(is + 1 ) (20)
where m is sufficiently large to guarantee that the IMC (( -,6 2J 0260) (25c)
controller q is proper. A is directly related to speed of DCi t(0 -,6+2/d)
the closed-loop response. The larger A make slower the The extra degree of freedom ,6 is calculated by solving
response and the smaller the actions of the manipulated [ Gq] 0 . That means we want to choose /B so
variable. With Eq. (20), the maximum peak for t f is 1;
that the term [1 -Gq] has a zero at the pole of GD.
i.e., the robustness characteristics are good.
For m > 1, filter forms other than Eq. (20) can lead to Therefore, we have [i (fli+1)eS (i225+1+=)] =0 -

faster response. For example, for m = 2, the filter The value of ,6 after some simplification is given as
f (1)({2S2 ++2A4s +1)
= (21) j T[(A2 + 2,T + T2 ) e/o
= 0/ 2A 2
The Eq. (21) is a standard second-order form, the
resonant peak in second order, Mp are uniquely related 2.4. Maximum sensitivity to modeling error
To evaluate the robustness of a control system, the
to the damping ratio J.
The relationship are given as maximum sensitivity M, which is defined by
MP I ,

22 ,[F$
for J<.0.707.
For J>0.707, there is no Ms = max I /[1 + GpGc (iw)] ,is used. Since the M is the

3326
inverse of the shortest distance from the Nyquist curve In FODUP process having gain uncertainty
of the loop transfer function to the critical point (-1,0),
A,, (s)
\
(K+AK) K AK (33)
a small value indicates a large stability margin of a K K
control system. For fair comparison under the same The process having gain uncertainty (Am (s)= AK/K),
robustness level, throughout all our simulation examples robust stability constraint should be
has been done i.e., all the controllers compared were
designed to have the same M values. co1222 +(2 0)2 ~ AKg (34)
3. ROBUST STABILITY r(2 + 2iis +_ 2) e01 /2 1}2 + I
j2 K

3.1 Norm-bound uncertainty regions Assuming FODUP process has uncertainty in time delay
Theorem: Robust Stability: (Morari and Zafirou, [6]). Am (S) = (e s
- 1) (35)
Assume that all plants p in the family HI Then for the delay uncertainty Eq. (32) should be
P( p (it)) (26) jr_2{(A7±2,Jr+ 2)e9 /Z-r2 }+2, 1 +° (36)

have the same number of RHP poles and that a

\(rl i 0 ) +(2 02;6) eAy-
Suppose that a small measurements time constant Ar
particular controller GC stabilizes the nominal plant p . was neglected in developing the nominal model. The
Then the system is robustly stable with the controller actual process transfer function is given as
GC if and only if the complementary sensitivity G,(s) =G /(Ars+l) and A (s) S)(A
= Az 1s-I Azs
z-s +1) (A z-s +1)
function 4 for the nominal plant p satisfies the
following bound A(jw) A c (37)
(ArT2aC/ +1I)
4 m sup m(i(Co) < 1 (27) Substituting the above result in Eq. (32), we have
For IMC controller (1- A22) +(2A4c0)2 ArC (38)
i = bq = 54f (28) Jr2 {(A2 + 2AJr + -2)9eIt /Z2 _ 1}2 02+1 (A 2 +1)
1(s)AAm (S) <l< (29) Process has uncertainty in all the three process
where Am (s) defines the process multiplicative parameters, 0t, , K
uncertainty bound. i.e., Am (s) =(Gp -6P)/P This I+ AK e- AOs
K
uncertainty bound can be utilized to represent the model A(s) K) (39)
(A r-s + 1)
reduction error, process input actuator uncertainty, and Substituting the above result in Eq. (36), we have
the process output sensor uncertainty etc, which is very
common in real process plant. (40)
For FODUP, the complementary sensitivity function 2
t{(2 +22Ar+ T)e I /T -1}2 +1 1
w,
t0>0
4 (s) is given for the proposed filter /(1i 22w2)2+(22;w)29 (A K
K
-1~
1 (s) (,3s + 1) e- /(22s2+ 22;s + 1)
= (30) (A1-jw+ 1)
where is f = r[( 2 + 2,1r+r2 ) eorl /T2 i2 The Eq. (40) is very useful to adjust the A if there are
P
Substituting value and Eq. (30) into Eq. (29) yields uncertainty in process parameters.
the robust stability constraint for tuning the adjustable 4. SIMULATION STUDY
parameters A .
Example 1. FODUP
-t{( A2 + 2A2z + -2 ) e"It'/Z/2 1} S + I2 1 (3 1)
_
Consider an unstable process, with process and
(2S2 +22As+1) Z\m
A (S)L disturbance models as follows [1, 2, 7, 10].
le-0.4s
Substituting s = j1) in above equation and we have Gp =GD=l (41)
In the recent, Liu et al. method [10] had already
Z'2 {(A2 + 2AJ- + -2 ) eol"/T/ 2
-
if 02 +I (32) demonstrated its superiority over many other previous
1
\1(
( A2a )2 (2Aj 1})w2+1J0) Am (S)_ approaches. The three controller parameters were taken
C(s)=(s+1)1(0.4s+l)
_

as Kc = 2 A = 0 = 0.4 , , and
There is possibility that uncertainty may exist in any of F(s) =2.89+ +0.469s' In order to achieve the same
the three process parameters i.e., 0, r, andK . 0.72s
Ms = 3.65 with Liu et al. [10], for the proposed method
Therefore, we have to consider the uncertainty in A = 0.401 and j = 0.72 has been adjusted and
different parameters separately.
corresponding tuning parameters settings are

3327
Kc =2.857,TI=1.759 and zD=0.152 . A unit step Example 2. SODUP
change is added to the setpoint input at t = 0 and an The following unstable process was considered for the
inverse unit step change of load disturbance is added to present study [1,2,7].
the process input at t = 4. The simulation results are G1 C
i -0.5s (i49'7
provided in Fig. 2. It is seen that the proposed method (50s 1)(2s+1)(O.5s+1)
results in the improved load disturbance response. The The above model has been approximated to the SODUP
IATE values are 0.791, 1.008 and 1.006 for disturbance by [1,2,7] and given as G =GD=le-0939(5s -1)(2.07s+1)
rejection for the proposed, Liu et al. [10] and Tan et al. The A.2 & =0.71 and i= 1.5 have been used for the
[7] method respectively. The setpoint response of the
Liu et al. [10] and Tan et al. [7] are smooth and simulation study for the proposed and Yang et al. [8]
overlapping. Figure 3 shows the controller output, method respectively, that has Ms = 2.216. For the Tan et
where Liu et al. [10] and Tan et al. [7] have sharp long al. [7] method the three controller parameters
peak and proposed method has smaller peak. The sharp are Ko = 2(2.07s + 1) K = s+l and K2= 3.58(2.4s + 1)-
,
peak in the controller output for the practical situation 0.2s +1
are undesirable, where there are inequality constraints For the simulation, the setpoint has a step change of
on the manipulated variable and controller saturation magnitude 1 at t=0 and the load disturbance has an
may happen. Although the setpoint for the Liu et al. [10] inverse step change of magnitude 1 at t=30, the time
and Tan et al. [7] have smooth response but due to response is shown in Fig. 4. The proposed tuning
saturation of controller output the above methods may method has a fast settling time compared to other
have worst performance with compare to the proposed existing methods. Tan et al. [7] has slow output
method for setpoint. response for both setpoint as well as disturbance
rejection. For the setpoint the 2DOF controller structure
is used and b=0.2 is selected for both proposed and
Yang et al. [8] method. It is clear that at the same
robustness level (Ms=2.216) the proposed method has
clear advantage in both setpoint and disturbance
,D 0.8-
Proposed
rejection.
- - - Liu et al.
a) 0.6 Tan et al.
a)

ry 0.64

,D 0.8
0.2

r 0.6
0
0 6
Time
o Proposed
L 0.4 Tan et al.
-Yang et al.
Fig. 2. Simulation results for Example 1 - -

2.5 0.2
Proposed
-Liuetal.
Tan et al. 0<
0 10 20 30 40 50 60
Time
1 .5
Fig. 4. Simulation results for Example 2
Example 3. SODUP with negative zero
0 Consider an unstable process with a strong lead time
constant and two unstable poles [2] as follows:
2 (5s +1) e

-0.5 (3s - 1) (ls - 1)

Lee et al. [2] and proposed methods were used to design
-1 the PID controller. For the proposed method, a value of
A=0.657 and S =1.6 was chosen so that Ms = 4.49.
-1.5
0 1 2 3 4 5 The A = 0.692 has been adjusted to get the same value
Time of the Ms for Lee et al. [2] to obtain the fair
Fig. 3. Setpoint controller output for Example 1
comparison. Figure 5 shows the closed-loop output
response for a unit-step setpoint change occurring at t=0,
and an inverse unit-step step change of load is added to

3328
the process input at t=15. Figure 5 shows that the ACKNOWLEDGEMENT
disturbance rejection and set-point response for the This research was supported by the BK21 program by
proposed controller is better than the Lee et al. [2] the Ministry of Education & Human Resources
tuning methods. The setpoint response of the proposed Development.
method is fast and settling time is less compare to above
said method, which has slow response and long settling REFERENCES
time. In the disturbance rejection, both overshoot and
undershoot are small in the proposed tuning method. [1] H. P. Huang and C. C. Chen, "Control-system
Lee et al. [2] tuning method show bigger peak and large synthesis for open-loop unstable process with time
undershoot, which is highly undesirable in the control delay," IEE Process-Control Theory and
system. Application, Vol. 144, pp. 334, 1997.
[2] Y. Lee, J. Lee and S. Park, "PID controller tuning
for integrating and unstable processes with time
3 delay," Chem. Eng. Sci., Vol. 55, pp. 3481-3493,
Proposed 2000.
Lee et al. [3] Y. G. Wang and W. J. Cai, "Advanced
proportional-integral-derivative tuning for
integrating and unstable processes with gain and
a) 1 phase margin specifications," Ind Eng. Chem. Res.
cn
Vol. 41, No. 12, pp. 2910-2914, 2002.
cn
[4] Y. Lee, S. Park, M. Lee and C. Brosilow, "PID
cn controller tuning for desired closed-loop responses
for SI/SO systems," AIChE Journal, Vol. 44, No. 1,
m
1
pp. 106-115, 1998.
[5] S. Majhi and D. P. Atherton, "Obtaining controller
-2-
parameters for a new Smith Predictor using
autotuning," Automatica, Vol. 36, pp. 1651-1658,
2000.
-3
0 5 10 15 20 25 30 [6] M. Morari and E. Zafiriou, "Robust Process
Time Control," Prentice-Hall: Englewood Cliffs, NJ,.
1989.
Fig. 5. Simulation results for Example 3 [7] W. Tan, H. J. Marquez and T. Chen, "IMC design
4.1 Selection of the tuning parameter ; for unstable processes with time delays," J Process
It is well-known that there is always a tradeoff in Control, Vol. 13, pp. 203-213, 2003.
selecting the desired closed-loop tuning parameter A. [8] X. P. Yang, Q. G. Wang, C. C. Hang and C. Lin,
Fast speed of response and good disturbance rejection "IMC-based control system design for unstable
are favored by choosing a small value of A. however, processes," Ind Eng. Chem. Res. Vol. 41, No. 17,
stability and robustness are favored by a large value of A. pp. 4288-4294, 2002.
Hence, the choice of i is entirely based on the [9] W. D. Zhang, D. Gu, W. Wang, and X. Xu,
experience of the operator with the control system. "Quantitative performance design of a modified
Based on many simulation studies, it is observed that Smith Predictor for unstable processes with time
the starting value of i can be considered to be slightly delay," Ind Eng. Chem. Res. Vol. 43, No. 1, pp.
more than the process time delay, which can gives 56-62, 2004.
robust control performances. If not, the value should be [10] T. Liu, W. Zhang and D. Gu, "Analytical design of
increased carefully until both the nominal and robust two-degree-of-freedom control scheme for open-
control performances are achieved. loop unstable process with time delay," J. Process
Control, Vol. 15, pp. 559-572, 2005.
5. CONCLUSIONS
The IMC filter structures have been modified for several
representative unstable processes to improve
disturbance rejection performance of the PID controller.
Based on the proposed filter structures, tuning rules for
the PID controller was derived by using the generalized
IMC-PID method by Lee et al. [4]. For the unstable
process with negative zero, undershoot in disturbance
rejection can be eliminated by the overdamped IMC
filter. The simulation results demonstrated superiority of
the proposed method for the same robustness level
(keeping same M s ) to other tuning methods.

3329