You are on page 1of 8

Chemical Engineering Science 63 (2008) 5141 -- 5148

Contents lists available at ScienceDirect

Chemical Engineering Science


journal homepage: w w w . e l s e v i e r . c o m / l o c a t e / c e s

Robust PID controller design for nonlinear processes using JITL technique
Cheng Cheng, Min-Sen Chiu ∗
Department of Chemical and Biomolecular Engineering, National University of Singapore, 4 Engineering Drive 4, Singapore 117576, Singapore

A R T I C L E I N F O A B S T R A C T

Article history: A robust PID controller design methodology for nonlinear processes is proposed based on the just-in-time
Received 14 May 2007 learning (JITL) technique. To do so, a composite model consisting of a nominal ARX model and the JITL,
Received in revised form 18 June 2008 where the former is used to capture linear process dynamics and the latter to approximate the inevitable
Accepted 10 July 2008
modeling error caused by the process nonlinearity, is employed to model the process dynamics in the
Available online 15 July 2008
operating space of interest. The state space realizations of this composite model and PID controller
Keywords:
are then reformulated as an uncertain closed-loop system, by which the corresponding robust stability
Robust control condition is developed. Literature examples are employed to illustrate the proposed methodology and a
Just-in-time learning comparison with the previous result is made.
Nonlinear system © 2008 Elsevier Ltd. All rights reserved.
PID controller

1. Introduction nonlinear CSTR for which a first-principle model is assumed to be


available. By assuming that the process/model mismatch is entirely
For most chemical processes, the first-principle models are usu- due to the nonlinearities of the process, bounds on the conic sectors
ally unavailable because of the lacking of physicochemical knowl- that describe the process nonlinearities were developed and used in
edge. An attractive alternative for controller design is to rely on the standard M −  structure for robust stability analysis. However,
the models extracted from process input and output measurements. the identification of the conic bounds is not trivial and the result-
These models generally have varying degrees of accuracy. If the ing robust stability analysis tends to give conservative result, not
plant/model mismatch is not taken into account in controller design, to mention that first-principle models are generally not available
the resulting control performance may become poor or even unsta- for many chemical processes. To lessen the modeling requirement,
ble in the presence of significant modeling error. This design problem Knapp and Budman (2000) developed a robust PID controller design
has motivated the researchers to pursue various robust controller methodology for nonlinear processes using an empirical state affine
design methods in the last two decades (Morari and Zafiriou, 1989; model developed from the available process data, which can be
Malan et al., 2004). The objective of robust controller design is to en- readily transformed into a suitable form for the robust stability anal-
sure closed-loop stability and maintain control performance not only ysis. Although their result provides an attractive alternative to the
for the nominal model but also for a set of possible process models Doyle's approach, the construction of the state affine models is rather
that captures the actual process dynamics. For linear system, vari- tedious and computational demanding. According to Knapp and
ous robust control problems had been tackled by using the transfer Budman (2000), in order to obtain a state affine model, a NARMA
function model approaches (Morari and Zafiriou, 1989; Packard and model is initially constructed from the available process input and
Doyle, 1993). Normally, a given set of process models is represented output data. Subsequently, an algorithm developed by Diaz and
by a nominal model together with a suitable uncertainty descrip- Desrochers (1988) is employed to find the parameters for a trun-
tion equation to account for the modeling error between the nomi- cated Volterra model based on the NARMA model identified previ-
nal model and actual process dynamics. The associated design issue ously. Once the Volterra kernels are obtained, a generalized Hankel
of estimating the uncertainty models also attracted much research matrix can be developed to find a state affine model (Sontag, 1979).
investigations (Gustafsson and Makila, 2001; Boling et al., 2004). Obviously, the modeling efforts required to identify a state affine
Some robustness analysis results developed for linear systems model are extensive and thus hampers the application of robust
have also been applied to the nonlinear systems. For example, controller design method developed based on such a model.
Doyle et al. (1989) proposed a robust controller design method for a To circumvent the aforementioned drawbacks by using the conic
sectors and state affine models in robust controller designs for non-
linear systems, a robust PID controller design methodology using the

Corresponding author. Tel.: +65 6516 2223; fax: +65 6779 1936. just-in-time learning (JITL) technique (Cybenko, 1996; Atkeson et al.,
E-mail address: checms@nus.edu.sg (M.-S. Chiu). 1997; Bontempi et al., 2001) is developed in this paper. It is noted

0009-2509/$ - see front matter © 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ces.2008.07.013
5142 C. Cheng, M.-S. Chiu / Chemical Engineering Science 63 (2008) 5141 -- 5148

that robust PID controller design methods have been developed us-
ing different techniques in the literature, for example linear matrix z−1
inequalities (LMIs) approach in the multiple-model modeling frame-
x(k) x(k + 1)
work (Ge et al., 2002; Toscano, 2007) and numerical optimization
approach by solving the maximization of the shortest distance from
the Nyquist curve of the open-loop transfer function to the critical A0 B0 E1
point (Toscano, 2005). However, these methods require a priori pro- u(k) C0 D0 E2 y(k)
cess knowledge or ad-hoc procedure to address the associated de-
F1 F2 E3
sign issue of partition of the operating space in the multiple-model
approach and determine the worst-case model (or optimal model for
that matter) in the operating space of interest in order to obtain the ~
y(k)
best viable PID design in the numerical optimization approach. In ~
u(k)
contrast, the proposed design method is a one-shot design approach 1
in the sense that the PID parameters are obtained directly from a 2
set of process data characterizing the process dynamics of interest.
3
In this respect, the proposed method is more advantageous than the
previous model-based robust PID design methods because the re- a
quired a priori process information may not be readily available in
practical applications, leading to the tedious trial and error design Fig. 1. The interconnection structure for the composite model described by Eqs. (8)
procedure. In the proposed method, it is assumed that the process and (10).
nonlinearity is the only source of the model uncertainty. A compos-
ite model consisting of a nominal ARX model and the JITL, where where x(k) = [x1 (k) x2 (k)]T and
the former is used to capture linear process dynamics and the latter  
is applied to approximate the inevitable modeling error caused by  + 1 l,2 + 2
A = l,1
the process nonlinearity. The state space realization of this compos- 1 0
ite model is then reformulated as an uncertain system, by which the
robust stability condition of this uncertain system under PID control B0 = [1 0]T ; C = [l,1 + 1 0]; D0 = 0
is developed. Literature examples are used to illustrate the proposed
Given the process input and output data {u(k), y(k)}, the
method and a comparison with the previous result is made.
plant/model mismatch caused by the nonlinearity, i.e. y(k) − yl (k),
can be calculated after the nominal ARX model given in Eq. (2)
2. Modeling methodology is identified. Subsequently, the JITL technique can be applied to
model the dynamics between the input sequence {u(k)} and the
In this paper, a composite model consisting of a nominal ARX sequence {y(k) − yl (k)} by using the reference dataset constructed
model and JITL models is used to describe the nonlinear process in from {u(k), y(k)}. Upon the successful implementation of JITL algo-
the operating range of interest, where the former can be identified rithm, the respective ranges of variation for model coefficients in
by using the process input and output data around a nominal oper- Eq. (3) are denoted by 1 ∈ [1,min 1,max ], 2 ∈ [2,min 2,max ],
ating condition and the latter is used to capture the modeling error and 1 ∈ [1,min 1,max ], which can be represented by the follow-
caused by the process nonlinearity, i.e. the difference between the ing equations:
predicted output of nominal ARX model and actual process output.
Suppose that an input sequence {u(k)} is injected into the process and 1 = ¯ 1 (1 + r1 1 ); |1 |  1
the corresponding output sequence {y(k)} is measured. The follow- 2 = ¯ 2 (1 + r2 2 ); |2 |  1
ing composite model is then used to model the nonlinear dynamics 1 = ¯ 1 (1 + r3 3 ); |3 |  1 (6)
between {u(k)} and {y(k)}:
where i (i = 1–3) is the uncertainty bounded by one and other
ŷ(k) = yl (k) + ynl (k) (1) parameters are defined as follows:
1,min + 1,max 1,max − 1,min
¯ 1 = ; r1 =
where ŷ(k) is the output of the composite model, yl (k) is the predicted 2 1,min + 1,max
output of nominal ARX model, and ynl (k) is the effect of process  + 2,max 2,max − 2,min
nonlinearity, i.e. y(k)−yl (k), to be approximated by the JITL technique ¯ 2 = 2,min ; r2 =
2 2,min + 2,max
(see Appendix A).
 +  1,max − 1,min
Because JITL normally employs a first-order or second-order ARX ¯ 1 = 1,min 1,max
; r3 = (7)
model, we shall use a second-order model structure for both nominal 2 1,min + 1,max
ARX model and JITL in the subsequent developments. Therefore, yl (k) By using Eqs. (6) and (7), Eqs. (4) and (5) can be rewritten as
and ynl (k) are represented as follows:      
A1 A2
x(k + 1) = A0 + 1 +  x(k) + B0 u(k) (8)
yl (k) = l,1 y(k − 1) + l,2 y(k − 2) + l,1 u(k − 1) (2) 01×2 01×2 2

ŷ(k) = (C0 + C1 3 )x(k) + D0 u(k) (9)


ynl (k) = 1 y(k − 1) + 2 y(k − 2) + 1 u(k − 1) (3)
where 0n×m denote a n × m zero matrix and
Consider the following state space realization of the composite  
model given in Eqs. (1)–(3):  + ¯ 1 l,2 + ¯ 2
A0 = l,1
1 0
x(k + 1) = Ax(k) + B0 u(k) (4) A1 = [¯ 1 r1 0]; A2 = [0 ¯ 2 r2 ]

ŷ(k) = Cx(k) + D0 u(k) (5) C0 = [l,1 + ¯ 1 0]; C1 = [¯ 1 r3 0]


C. Cheng, M.-S. Chiu / Chemical Engineering Science 63 (2008) 5141 -- 5148 5143

To account for the modeling error resulting from the approxi-


mation of the nonlinear process by the proposed composite model,
z−1 an additive uncertainty a is appended to the model output ŷ(k) as
x(k) x(k + 1) follows:
Ψ(k) Ψ(k + 1)
y(k) = ŷ(k) + la x1 (k)a , |a |  1 (10)
M11 M12
where la is the magnitude of the worst perturbation as calculated by
M21 M22  
 y(k) − ŷ(k) 
 
la = Max   (11)
k  ŷ(k)/(l,1 + 1 ) 
~ (k) ~
y(k)
u
1
2
3
a 1

0.9
M
0.8

0.7
M11 M12
0.6
M21 M22 y
x(k) x(k + 1) 0.5
Ψ(k) Ψ(k + 1)
~ ~
y (k) 0.4
u(k)
zI4
0.3 Process output
1 ARX model
2 Composite model
0.2
3 0 50 100 150 200 250 300
a Sample

Fig. 2. The interconnection structure for the uncertain closed-loop system. Fig. 4. Validation result for the CSTR example.

0.5
y

0
0 50 100 150 200 250 300

0.5
yl

0
0 50 100 150 200 250 300

0.2
y-yl

-0.2
0 50 100 150 200 250 300

30

15
u

0
0 50 100 150 200 250 300
Sample
Fig. 3. Input–output data used for the JITL.
5144 C. Cheng, M.-S. Chiu / Chemical Engineering Science 63 (2008) 5141 -- 5148

It is noted from Eq. (9) that the state variable x2 (k) has no effect on ỹ(k) = [¯ 1 r1 x1 (k) ¯ 2 r2 x2 (k) ¯ 1 r3 x1 (k) la x1 (k)]T
the model output ŷ(k), which explains why the additive uncertainty
a in Eq. (10) is only associated with the state variable x1 (k). ũ(k) = [¯ 1 r1 x1 (k)1 ¯ 2 r2 x2 (k)2 ¯ 1 r3 x1 (k)3 la x1 (k)a ]T
After some algebraic manipulation, the composite model by in-
cluding the uncertainty a , i.e. Eqs. (8) and (10), can be recast into Lastly, when a first-order ARX model is employed for both nomi-
the interconnection structure as depicted in Fig. 1, where nal model and the JITL, the corresponding interconnection structure
is given by
x(k + 1) = A0 x(k) + B0 u(k) + E1 ũ(k) (12)
A0 = l,1 + ¯ 1 ; B0 = 1
y(k) = C0 x(k) + D0 u(k) + E2 ũ(k) (13)
C0 = l,1 + ¯ 1 ; D0 = 0
ỹ(k) = F1 x(k) + F2 u(k) + E3 ũ(k) (14)
E1 = [1 0 0]; E2 = [0 1 1]; E3 = 03×3
 
1 1 0 0
E1 = ; E2 = [0 0 1 1]; E3 = 04×4 F1 = [¯ 1 r1 ¯ 1 r3 la ]T ; F2 = 03×1
0 0 0 0
⎡ ⎤
¯ 1 r1 0 ỹ(k) = [¯ 1 r1 x(k) ¯ 1 r3 x(k) la x(k)]T
⎢ 0 ¯ 2 r2 ⎥
F1 = ⎢ ⎥; F2 = 04×1
⎣ ¯ r
1 3 0 ⎦ ũ(k) = [¯ 1 r1 x(k)1 ¯ 1 r3 x(k)3 la x(k)a ]T
la 0
3. Robust stability analysis

Since PID controller is widely used in the process industries, it


is considered in the ensuing robust stability analysis. To do so, PID
controller is represented by the following state space equation:

W(k + 1) = Ac W(k) + Bc e(k) (15)

u(k) = Cc W(k) + Dc e(k) (16)

where W(k) is a 2 × 1 state variable vector of PID controller, e(k)


is the tracking error, i.e. the difference between the set-point and
process output, and other model parameters are given by
 
1 0 1
Ac = ; Bc =
1 0 0
 
kc kc
Cc = − kc D kc D ; Dc = kc + + kc D (17)
I I
where kc , I , and D are the PID parameters.
By using Eqs. (12)–(16), the uncertain closed-loop system under
Fig. 5. Robust stability region (shadow) for the CSTR example. PID control as depicted in Fig. 2(a) is represented by the following

0.8

0.6
x1

0.4

0.2

0
0 20 40 60 80 100 120

1
Set-point
0.8 Process output
0.6
x1

0.4

0.2

0
0 20 40 60 80 100 120
Sample
Fig. 6. Set-point responses for the PI designs with I = 1.1 and (1) kc = 32.4 (top); (2) kc = 50 (bottom).
C. Cheng, M.-S. Chiu / Chemical Engineering Science 63 (2008) 5141 -- 5148 5145

augmented state space equation: 0.57


   
x(k + 1) x(k)
= M11 + M12 ũ(k) (18)
W(k + 1) W(k) 0.56
 
x(k)
ỹ(k) = M21 + M22 ũ(k) (19)
W(k) 0.55
where the matrices M11 , M12 , M21 , and M22 are given by
  0.54

x1
A − B0 Dc C0 B0 Cc
M11 = 0
−Bc C0 Ac
  0.53
E − B0 Dc E2
M12 = 1 Set-point
−Bc E2 Process output
0.52
M21 = [F1 04×2 ]

M22 = 04×4 0.51


0 10 20 30 40 50 60 70 80
In Fig. 2(a), the delay operator z−1 can be considered as a pertur- Sample
bation bounded by one and therefore can be treated as an artificial
uncertainty z with |z |  1, resulting in a repeated perturbation z I4
(I4 standing for a 4 × 4 identity matrix) in Fig. 2(b), where the over- 0.88
all uncertainty block is given by  = diag[z I4 1 2 3 a ]. Conse-
0.875
quently, the robust stability condition of this uncertain closed-loop
system can be obtained by using the structured singular value test 0.87
(Packard and Doyle, 1993) as follows:
  0.865
M11 M12
 <1 (20)
M21 M22 0.86
x1

where  denotes the structured singular value with perturbation 0.855


structure .
Lastly, when the composite model is constructed by a first-order 0.85
ARX model and a PI controller is desired, Eq. (20) can be applied to 0.845
by using the relevant model parameters provided at the end of last
section and the following controller parameters, Ac = 1, Bc = 1, Cc = 0.84
kc / I , and Dc = kc + kc / I , along with the corresponding perturbation
block given by diag[z I2 1 3 a ]. 0.835
0 10 20 30 40 50 60 70 80
4. Examples Sample

Example 1. Consider a CSTR described by the following equations 0.208


(Doyle et al., 1989):
0.206
ẋ1 = −x1 + Da(1 − x1 ) ex2 /(1+x2 / ) (21) 0.204

ẋ2 = −x2 − b(x2 − xc ) + BDa(1 − x1 ) ex2 /(1+x2 / ) (22) 0.202

where x1 and x2 are the dimensionless concentration and temper- 0.2


ature of the reactor, and xc (=u) is the cooling temperature chosen
x1

0.198
as manipulated variable while x1 (=y) is the controlled variable. The
process has one stable steady state when Da = 0.072, B = 1, b = 0.3, 0.196
and  = 20. The following operating space xc ∈ [5 23] and x1 ∈ 0.194
[0.1969 0.8781] is considered in this example.
0.192
To construct the composite model, a first-order ARX model is 0.19
adopted. By using the process input and output data obtained around
the nominal operating condition, the parameters of the nominal ARX 0.188
model are obtained as l,1 = 0.7216 and l,1 = 0.1231. To model
0 10 20 30 40 50 60 70 80
the process nonlinearity by the JITL, a different set of input and Sample
output data is generated within the operating space as illustrated in
Fig. 7. Closed-loop responses for −20% step disturbance at three operating conditions.
Fig. 3, where yl is the predicted output by the nominal ARX model
and y − yl is the modeling error caused by the process nonlinearity.
To model the nonlinearity effect by the JITL with = 0.75, kmin =
10, and kmax = 60, the resulting model parameters are obtained as
1 ∈ [−0.3363, 0.0191] and 1 ∈ [−0.0217, 0.1112]. In addition, resulting composite model gives reasonably good prediction in the
to quantify the modeling error of the composite model, the worst validation test by using input and output data different from that
perturbation is calculated as la = 0.0514. Fig. 4 illustrates that the used in constructing the composite model.
5146 C. Cheng, M.-S. Chiu / Chemical Engineering Science 63 (2008) 5141 -- 5148

0
-0.05

y
-0.1
0 100 200 300 400 500 600 700 800

0
yl -0.05
-0.1
0 100 200 300 400 500 600 700 800

x 10-3
5
0
y-yl

-5
-10
0 100 200 300 400 500 600 700 800

0.05
0
u

-0.05
0 100 200 300 400 500 600 700 800
Sample

Fig. 8. Input–output data used for the JITL.

0.01

-0.01

-0.02

-0.03
y

-0.04

-0.05

-0.06
Process Output
-0.07 Nominal ARX Model
Composite Model
-0.08
0 50 100 150 200 250 300 350 400 450 500
Sample

Fig. 9. Validation results for the distillation process. Fig. 10. Robust stability region (shadow) for the distillation process.

To proceed with the proposed robust PI design, Eq. (20) is ap- this PI controller yields stable response at any operating condition
plied to obtain the PI parameters that guarantee the robust stability within the pre-specified operating space.
over the concerned operating space. Fig. 5 shows the resulting ro- Lastly, it is noted that this example was previously studied by
bust stability region in the kc − I parameter space. To verify the Knapp and Budman (2000) who developed a robust PID design
analysis result thus obtained, set-point changes covering the entire method based on the state affine model. Compared with the viable
operating space for the viable PI design with kc = 32.4 and I = 1.1 PI designs reported in their paper, the proposed method gives less
are conducted. As can be seen from Fig. 6, this PI controller is able to conservative result, for example, the maximum allowable kc ob-
maintain closed-loop stability over the entire operating space. For PI tained for I = 1.1 is 18 in their result, not to mention that tedious
design with I = 1.1, the actual maximum allowable controller gain modeling procedure is required to build a state affine model.
is found to be kc = 50 through the trial and error procedure via dy-
namic simulation studies. The set-point responses of this controller Example 2. Consider a distillation process described by (Gao et al.,
are also shown in Fig. 6 for comparison purpose. To illustrate the 2000):
disturbance rejection capability of the PI controllers thus obtained,
−20% step disturbance in the parameter b is considered. The con- y(k) = 0.757y(k − 1) + 0.243g(u(k − 1)) (23)
trol performances of the above PI design (kc = 32.4 and I = 1.1) at
different operating conditions are illustrated in Fig. 7. It is clear that g(x) = 1.04x − 14.11x2 − 16.72x3 + 562.7x4 (24)
C. Cheng, M.-S. Chiu / Chemical Engineering Science 63 (2008) 5141 -- 5148 5147

0.02 systems. It is also known as instance-based learning (Aha et al., 1991),


Set-point local weighted model (Atkeson et al., 1997), lazy learning (Aha, 1997;
Process output Bontempi et al., 2001), or model-on-demand (Braun et al., 2001) in
0
the literature. The JITL has no standard learning phase because it
merely stores the data in the database and the computation is not
-0.02 performed until a query data arrival by constructing a local model
to approximate the process dynamics characterized by the current
query data. The JITL algorithm adopted in this paper is now briefly
-0.04
y

reviewed in the sequel.


Suppose that a database consisting of N process data (yi , xi )i=1∼N ,
-0.06 yi ∈ R, xi ∈ Rn , is collected. Given a specific query data xq ∈ Rn , the
objective of JITL is to predict the model output ŷq = f (xq ) according
to the known database (yi , xi )i=1∼N . To do so, the relevant data are
-0.08 selected from the database first by using the following similarity
measure si (Cheng and Chiu, 2004):
-0.1 
2
0 20 40 60 80 100 si = e−d (xq ,xi ) + (1 − ) cos(
i ) if cos(
i )  0 (25)
Sample
where is a weight parameter and is constrained between 0 and 1,
Fig. 11. Set-point response for the PI design with kc = 1.24 and I = 2.5. and
i is the angle between xq and xi , where xq = xq − xq−1 and
xi = xi − xi−1 .
To apply JITL for the modeling of dynamic systems, all si are com-
where the top column composition y(%) is the process output and puted by Eq. (25) first and for each l ∈ [kmin kmax ], where kmin and
the reflux flow rate u (mol/min) is the process input. The operating kmax are the pre-specified minimum and maximum number of rele-
space under consideration is given by u ∈ [−0.05 0.01] and y ∈ vant data, the relevant data set (yl , l ), where yl ∈ Rl×1 and l ∈ Rl×n ,
[−0.0817 0.009]. Again, input and output data around the nominal are constructed by selecting l most relevant data (yi , xi ) correspond-
operating condition are used to identify the nominal ARX model with ing to the largest si to the l-th largest si . Denote Wl ∈ Rl×l , a diagonal
parameters obtained as l,1 = 0.7686 and l,1 = 0.2183. To apply the weight matrix with diagonal elements being the first l largest values
JITL to model the process nonlinearity, a different set of input and of si . Next, the leave-one-out cross-validation test is conducted and
output data as depicted in Fig. 8 is generated within the operating the validation error is calculated. Upon the completion of the above
space. With the parameters of JITL chosen as = 0.95, kmin = 8, and procedure, the optimal l, l∗ , is determined by that giving the smallest
kmax = 60, the resulting model parameters are obtained as 1 ∈ validation error. Subsequently, the predicted output for query data is
[−0.3750 0.0480] and 1 ∈ [−0.0133 0.6614]. In addition, the worst calculated as xqT (PlT∗ Pl∗ )−1 PlT∗ Wl∗ yl∗ , where Pl∗ = Wl∗ l∗ and Wl∗ is a
modeling error of the composite model is calculated as la = 0.0543. diagonal matrix with entries being the first Wl ∈ Rl×l largest si , pro-
To illustrate the predictive performance of nominal ARX model and vided this optimal model satisfies the stability constraint; otherwise,
composite model, Fig. 9 shows that the composite model gives better an optimization procedure is carried out to recompute the optimal
prediction than the nominal ARX model over the entire operating model.
space.

Fig. 10 gives the viable PI designs in the kc − I parameter space References


that meet the robust stability condition, Eq. (20). To verify the result
thus obtained, Fig. 11 shows the set-point responses of the PI con- Aha, D.W., 1997. Lazy learning. Artificial Intelligence Review 11, 7–10.
Aha, D.W., Kibler, D., Albert, M.K., 1991. Instance-based learning algorithms. Machine
troller with kc = 1.24 and I = 2.5. It is evident that this controller Learning 6, 37–66.
gives stable response for the set-point changes covering the entire Atkeson, C.G., Moore, A.W., Schaal, S., 1997. Locally weighted learning. Artificial
operating space. Intelligence Review 11, 11–73.
Boling, J.M., Haggblom, K.E., Nystrom, R.H., 2004. Multivariable uncertainty
estimation based on multi-model output matching. Journal of Process Control
5. Conclusion 14, 293–304.
Bontempi, G., Bersini, H., Birattari, M., 2001. The local paradigm for modeling and
control: from neuro-fuzzy to lazy learning. Fuzzy Sets and Systems 121, 59–72.
A new methodology for robust PID controller design of nonlinear
Braun, M.W., Rivera, D.E., Stenman, A., 2001. A model-on-demand identification
processes is developed. In the proposed method, a composite model methodology for nonlinear process systems. International Journal of Control 74,
is first constructed to model the process dynamics for the operating 1708–1717.
space of interest. This composite model consists of a nominal ARX Cheng, C., Chiu, M.S., 2004. New data-based methodology for nonlinear process
modeling. Chemical Engineering Science 59, 2801–2810.
model to capture linear process dynamics and the JITL to approxi-
Cybenko, G., 1996. Just-in-time learning and estimation. In: Bittanti, S., Picci, G.
mate the modeling error caused by process nonlinearity. The state (Eds.), Identification, Adaptation, Learning: The Science of Learning Models from
space realizations of this composite model and PID controller are Data. Springer, Berlin, pp. 423–434.
then reformulated as an uncertain closed-loop system, by which the Diaz, H., Desrochers, A., 1988. Modeling of nonlinear discrete-time system from
input/output data. Automatica 24, 629–641.
corresponding robust stability condition is developed. Simulation re- Doyle, F.J., Packard, A.K., Morari, M., 1989. Robust controller design for a nonlinear
sults show that the proposed method can be used to design robust CSTR. Chemical Engineering Science 44, 1929–1947.
PID controller to ensure the closed-loop stability for the concerned Gao, F., Wang, F., Li, M., 2000. An analytical predictive control law for a
class of nonlinear process. Industrial and Engineering Chemistry Research 39,
operating space.
2029–2034.
Ge, M., Chiu, M.S., Wang, Q.G., 2002. Robust PID controller design via LMI approach.
Appendix A. Just-in-time learning Journal of Process Control 12, 3–13.
Gustafsson, T.K., Makila, P.M., 2001. Modelling of uncertain systems with application
to robust process control. Journal of Process Control 11, 251–264.
The just-in-time learning (JITL) (Cybenko, 1996) technique was Knapp, T.D., Budman, H.M., 2000. Robust control design of non-linear processes using
developed as an attractive alternative for modeling the nonlinear empirical state affine models. International Journal of Control 73, 1525–1535.
5148 C. Cheng, M.-S. Chiu / Chemical Engineering Science 63 (2008) 5141 -- 5148

Malan, S., Milanese, M., Regruto, D., Taragna, M., 2004. Robust control from data Sontag, E., 1979. Realization theory of discrete-time nonlinear systems: Part I. The
via uncertainty model sets identification. International Journal of Robust and bounded case. IEEE Transactions on Circuits and Systems CAS-26, 324–356.
Nonlinear Control 14, 945–957. Toscano, R., 2005. A simple robust PI/PID controller design via numerical optimization
Morari, M., Zafiriou, E., 1989. Robust Process Control. Prentice-Hall, Englewood Cliffs, approach. Journal of Process Control 15, 81–88.
NJ. Toscano, R., 2007. Robust synthesis of a PID controller by uncertain multi-model
Packard, A., Doyle, F.J., 1993. The complex structured singular value. Automatica 29, approach. Information Science 177, 1441–1451.
71–109.

You might also like