You are on page 1of 18

ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Research article

Adaptive MPC based on MIMO ARX-Laguerre model


Imen Ben Abdelwahed, Abdelkader Mbarek n, Kais Bouzrara
Research Laboratory of Automatic Signal and Image Processing, National School of Engineers of Monastir, University of Monastir, 5019, Tunisia

art ic l e i nf o a b s t r a c t

Article history: This paper proposes a method for synthesizing an adaptive predictive controller using a reduced com-
Received 16 April 2016 plexity model. This latter is given by the projection of the ARX model on Laguerre bases. The resulting
Received in revised form model is entitled MIMO ARX-Laguerre and it is characterized by an easy recursive representation. The
12 November 2016
adaptive predictive control law is computed based on multi-step-ahead finite-element predictors,
Accepted 24 November 2016
identified directly from experimental input/output data. The model is tuned in each iteration by an
online identification algorithms of both model parameters and Laguerre poles. The proposed approach
Keywords: avoids time consuming numerical optimization algorithms associated with most common linear pre-
MIMO system modeling dictive control strategies, which makes it suitable for real-time implementation. The method is used to
ARX-Laguerre model
synthesize and test in numerical simulations adaptive predictive controllers for the CSTR process
Pole optimization
benchmark.
Online identification
Adaptive predictive control & 2016 ISA. Published by Elsevier Ltd. All rights reserved.

1. Introduction conventional approaches (PID, RST…), can be applied on different


system classes, such as Single-Input Single-Output (SISO) systems
The control of dynamical systems in the presence of large un- [1,6], nonlinear systems [4,8] and Multi-Input Multi-Output
certainties and constraints is of great interest for many applica- (MIMO) systems [9,10]. Other researchers have adopted the use of
tions. These problems occur when there are large variations of the other orthogonal bases in the case of Linear Time Invariant (LTI)
parameters due to changes in the operating region. To overcome and stable systems [11,12] or in the case of nonlinear systems
this problem an adaptive predictive control can be used [1–3]. To [13,14]. For a MIMO system, the developed approach consists in
formulate the adaptive model based predictive control, it is ne- decomposing the MIMO transfer matrix on the Generalized Or-
cessary to couple an online parametric optimization procedure thogonal Bases (GOB) [11]. The resulting model, known as MIMO-
with the predictive control algorithm. The synthesis of the Adap- GOB model, has many advantages with respect to classical models
tive Model Based Predictive Control is based on the use of an ex- (MIMO ARX model,…) such as insensitivity to the choice of sam-
plicit model. This Latter is used in a performance criterion or a cost pling time, the non-necessity of prior knowledge of the system
function to generate control signals by solving an optimization delay and particularly the parameter number reduction. It should
problem with respect to future control sequence leading the future be noted that the MIMO-GOB model is constrained by the diffi-
process outputs to desired values over a prediction horizon under culty of obtaining the optimal pole values of the GOB bases and by
certain constraints [4–7]. So the complexity of an adaptive pre- its restriction to decoupled MIMO systems. To overcome these
dictive control strategy depends on the of mathematical model disadvantages, Mbarek et al. [15] propose to decompose the MIMO
complexity. Consequently, for reducing this complexity a projec- ARX model coefficients on Laguerre orthonormal bases. The re-
tion of high order ARX models on orthonormal Laguerre bases is sulting model entitled MIMO ARX-Laguerre model ensures the
used for modeling complex systems. parameter number reduction with a recursive and easy re-
In recent years, some authors have adopted Laguerre functions presentation. An optimal choice of Laguerre poles is compulsory to
in order to reduce the complexity of the complex industrial pro- lessen considerably the number of MIMO ARX model parameters.
cess model and therefore, the predictive control law. The resulting We propose then an analytical solution for MIMO ARX-Laguerre
controller, boasted on its performances with respect to poles optimization by extending the works of [16]. The proposed
optimization consists in minimising the upper bound of the
n
squared norm of the error resulting from the truncation of the
Corresponding author.
series into a finite number of Laguerre functions. The MIMO ARX-
E-mail addresses: benabdelwahed.imen@gmail.com (I. Ben Abdelwahed),
abdelkader.mbarek@enim.rnu.tn (A. Mbarek), Laguerre poles are obtained using the knowledge of the MIMO
kais.bouzrara@enim.rnu.tn (K. Bouzrara). ARX-Laguerre parameters which are estimated recursively by

http://dx.doi.org/10.1016/j.isatra.2016.11.017
0019-0578/& 2016 ISA. Published by Elsevier Ltd. All rights reserved.

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
2 I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎

minimising the regularised square error to identify in offline To overcome this problem, a projection on the Laguerre bases is
manner the MIMO ARXLaguerre model. In the online case this proposed by Mbarek et al. [15]. In the following we detail the
procedure is processed using a sliding window where an online development of the MIMO ARX model on independent Laguerre
identification algorithm of the MIMO ARX-Laguerre model is ela- orthonormal bases to provide a reduced parameter number model.
borate. By applying the Predictive controller algorithm at each Due to stability condition in the meaning of Bounded Input
iteration of online identification, we obtain the Adaptive Predictive Bounded Output (BIBO) criterion, the coefficients hair and hbit sa-
Controller for MIMO process using the MIMO ARX-Laguerre model. tisfy:
The main contributions of this article is mainly threefold. (1) We ∞ ∞
propose an iterative algorithm to optimize the MIMO ARX-La- ∑ |hair (j)| < ∞; ∑ |hbit (j)| < ∞
guerre poles based on an analytical solution which depends on j=1 j=1 (6)
Fourier coefficients to guarantee the parameter number reduction such that
of this latter model. Furthermore, we propose a recursive algo-
rithm of Fourier coefficients using the regularized square error. ⎧
⎪ h a (j ) = 0 if j > nair
ir

(2) We propose the online identification algorithms of the MIMO ⎪
⎩ hbit (j ) = 0 if j > nbit (7)
ARX-Laguerre model. (3) By combining the online identification
parameters (Poles and Fourier coefficients) and the model based Then the coefficients hair and hbit are summable, they belong to
predictive control, we propose an Adaptive Predictive Control for Lebesgue space and since the Laguerre orthonormal functions
MIMO systems (APC MIMO) by using the MIMO ARX-Laguerre form an orthonormal basis in such space, the coefficients hair and
model. hbit can be decomposed on independent Laguerre bases with re-
This article is organised as follows: in Section 2, we present the spect to the outputs and the inputs respectively [17,18] as:
reduced complexity of ARX MIMO model by developing the ⎧ ∞
coefficients of the last model on independent Laguerre bases ⎪ hair (j ) = ∑ gn, air ℓ anir (j, ξair )
where we give a recursive representation of the MIMO ARX-La- ⎪ n= 0
⎨ ∞
guerre model. Section 3 deals with online parametric estimation ⎪
where we give a recursive identification of Fourier coefficients and ⎪ hbit (j ) = ∑ gn, bit ℓ bnit (j, ξ bit )
⎩ n= 0 (8)
an online poles optimization of the MIMO ARX-Laguerre model. In
Section 4, we present an Adaptive model based predictive control where ℓ anir (j,
ξair ) and ℓ bnit (j,
ξbit ) are the Laguerre functions and gn, air
of MIMO systems using the MIMO ARX-Laguerre model. Finally, and gn, bit are the Fourier coefficients.
Section 5 evaluates, through two simulation examples, the online By substituting relation (8) in each MISO ARX model given by
parametric identification of the MIMO ARX-Laguerre model and equation (2) we obtain the MIMO ARX-Laguerre where each MISO
the adaptive predictive control algorithm. output is defined as follows:
m Nair − 1 p N b it − 1
yi (k ) = ∑ ∑ gn, air x ni , yr (k ) + ∑ ∑ gn, bit x ni , ut (k )
2. MIMO ARX-Laguerre model r= 1 n= 0 t = 1 n= 0 (9)

where Nair and Nbit , (i, r = 1 ,.., m) ; (t = 1 ,.., p) are the truncated
A strictly causal discrete time MIMO system with p inputs and
orders of the MIMO ARX-Laguerre, gn, air and gn, bit represent the
m outputs can be described by a MIMO ARX model which can be
decomposed on a set of Multi-Input Single Output (MISO) system Fourier coefficients and x ni , yr and x ni , yu are the filtered outputs and
t
where the ith output (i = 1 ,.., m) is defined as: the filtered inputs, respectively, by Laguerre functions such as:
m nair p nb it ⎧
⎪ x ni , yr (k ) = ℓ anir (j, ξair )⁎ yr (k )
yi (k ) = ∑ ∑ hair (j) yr (k − j) + ∑ ∑ hbit (j) ut (k − j) ⎨
r=1 j=1 t=1 j=1 (1) ⎪ i b
⎩ x n, ut (k ) = ℓ nit (j, ξ bit )⁎ ut (k ) (10)
where ut(k) and yi(k) are the system inputs and outputs for where n is the convolution product.
(i, r = 1 ,.., m) and (t = 1 ,.., p), the nair and nbit are the model orders The Z-transform of relation (9) gives
associated to the outputs and the inputs respectively and hair and
m Nair − 1 p N b it − 1
hbit are the model parameters.
Yi (z ) = ∑ ∑ gn, air X ni , yr (z ) + ∑ ∑ gn, bit X ni , ut (z )
The ith output of the MIMO ARX model can be written in re- r= 1 n= 0 t = 1 n= 0 (11)
gression form as:
with Xni , yr (z ) and Xni , ut (z ) are the Z-transform of x ni , yr (k ) and x ni , ut (k )
yi (k ) = φT θ i (2) respectively.
with From relation (10), we can write:

⎪ X ni , yr (z ) = L nair (z ) Yr (z )
φ = ⎡⎣ y1 (k − 1), …, y1 (k − nai1 ), …, ym (k − 1), …, ym (k − naim ), ⎨
⎪ i
⎩ X n, ut (z ) = L nbit (z ) Ut (z )
u1 (k − 1), …, u1 (k − nbi1 ), …,up (k − 1), …, up (k − nbip ) ⎤⎦
T
(12)
(3)
where L nair (z ) and L nbit (z ) are the Z-transform of the Laguerre filters
θ i = ⎡⎣ hai1 (1), …, hai1 (nai1 ), …, haim (1), …, haim (naim ), given by relation (13) and (14), respectively:

hbi1 (1), …, hbi1 (nbi1 ), …, hbip (1), …, hbip (nbip ) ⎤⎦
T
(4) ⎪ air 1 − ξa2ir
⎪ L 0 (z ) =
z − ξair
This model is characterized by a high parameters number pn. ⎨ , n = 1, 2, …
⎪ 1 − ξair z air
m ⎛ m p ⎞ ⎪ L nair (z ) = L n − 1 (z, ξair )
⎩ z − ξair
pn = ∑ ⎜⎜ ∑ nair + ∑ nbit ⎟⎟ (13)
i=1 ⎝ r=1 t=1 ⎠ (5)

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 3

⎧ buit = 1 − ξ b2it ⎡⎣ 1 − ξ bit (ξ bit )2 ⋯ ( − ξ bit ) Nb it − 1⎤⎦


T
⎪ bit 1 − ξ b2it (22)
L
⎪ 0 (z ) =
z − ξ bit
⎨ , n = 1, 2, … Defining the state vectors X i (k ) of the ith output of the MIMO ARX-
⎪ 1 − ξ bit z bit
⎪ L nbit (z ) = L n − 1 (z, ξ bit ) Laguerre model as:
⎩ z − ξ bit (14)
X i (k ) = ⎡⎣ (X iy1 (k ))T ⋯ (X iym (k ))T (X ui1 (k ))T ⋯ (X uip (k ))T ⎤⎦
T
(23)
for (i, r = 1, … , m) and (t = 1, … , p).
By combining the recurrent form (13) and (14) of Laguerre or- and the inputs and the outputs vectors u (k ), y (k ) as:
thonormal functions with relation (12), we can formulate the re-
u (k ) = ⎡⎣ u1 (k ) ⋯ up (k ) ⎤⎦ , y (k ) = ⎡⎣ y1 (k ) ⋯ ym (k ) ⎤⎦
T T
currence relation for the filters Xni , yr (z ) and Xni , ut (z ), i = 1, 2, …, as: (24)
⎧ The ith output system is described by the following state re-
⎪ i 1 − ξa2ir
⎪ X 0, yr
(z ) = Yr (z ) presentation:
z − ξair

⎪ 1 − ξair z i ⎧ i i i i i
⎪ X (k + 1) = A X (k ) + Bu u (k ) + B y y (k )
⎪ X ni , yr (z ) = X n − 1, yr (z ) ⎨
⎩ z − ξair (15) ⎪ T
⎩ yi (k ) = Ci X i (k ) (25)

where Ci, (i = 1, … , m) is the parameter vectors containing the


⎧ Fourier coefficients gn, air for (n = 0, … , Nair − 1) and (r = 1, … , m)
⎪ i 1 − ξ b2it
⎪ X 0, ut (z ) = z − ξ Ut (z ) and gn, bit for (n = 0, … , Nbit − 1) and (t = 1, … , p):
⎨ bit
⎪ 1 − ξ bit z i ⎡
Ci = ⎣ g0, ai1 ⋯ g Na − 1, ai1 ⋯ g0, aim ⋯ g Na − 1, aim
⎪ X ni , ut (z ) = X n − 1, ut (z ) i1 im
⎩ z − ξ bit (16) ⎤T
g0, bi1 ⋯ g Nb ⋯ g0, bip ⋯ g Nb ⎥
i1− 1, bi1 ip − 1, bip ⎦ (26)
Then, the filtered outputs and the filtered inputs of the MIMO
ARX-Laguerre model can be written in recursive form as: with
⎧ i ⎡ Ai 0 ⎤
i
⎪ X yr (k + 1) = A iyr X iyr (k ) + b yr yr (k ); r = 1, …, m ⋯ 0 0 0
⎨ ; i = 1, …, m ⎢ y1 ⎥
⎪ X ui (k + 1) = A ui t X ui t (k ) +
i
but ut (k ); t = 1, …, p ⎢ ⋮ ⋱ ⋮ ⋮ ⋮ ⋮ ⎥
⎩ t (17)
⎢ ⎥
⎢ 0 ⋯ A iym 0 ⋯ 0 ⎥
with i
A =⎢ ⎥,
⎢ 0 ⋯ 0 A ui1 ⋯ 0 ⎥
⎧ ⎡ i ⎤T
⎪ X iy = i i
⎢⎣ x 0, yr (k ) x1, yr (k ) ⋯ x Na − 1, yr (k ) ⎥⎦ , (r = 1, …, m)
⎢ ⋮ ⋮ ⋮ ⋮ ⋱ ⋮ ⎥
⎪ r

ir ⎢ ⎥
⎪ i ⎡ i ⎤T
, (i = 1, …, m)
⎢⎣ 0 ⋯ 0 0 ⋯ A uip ⎥⎦
i (k ) ⋯ x Ni − 1, u (k ) ⎥ , (t = 1, …, p)
⎪ X ut = ⎢⎣ x 0, ut (k ) x1, ut ⎦
⎩ b it t (18) ⎡0 i ⎤ ⎡ Bi ′ ⎤
Bui = ⎢ ⎥, Biy = ⎢
pny × p y ⎥
where A yir
and Auit
for (i, r = 1, … , m) and (t = 1, … , p) are square ⎢ B i′ ⎥ ⎢0 i ⎥
⎣ u ⎦ ⎣ pnu × m ⎦ (27)
matrices of dimensions Nair and Nbit respectively, defined as fol-
lows: where

⎡ ξair 0 0 ⋯ 0⎤
⎢ ⎥
⎢ 1 − ξa2ir ξair 0 ⋯ 0⎥
⎢ ⎥
A iyr =⎢ 2
− ξair (1 − ξair ) 1 − ξair2
ξair ⋯ 0⎥
⎢ ⎥
⎢ ⋮ ⋮ ⋮ ⋱ ⋮ ⎥
⎢⎣ ( − ξa ) Nair − 2 (1 − ξa2 ) ( − ξa ) Nair − 3 (1 − ξa2 ) ( − ξa ) Nair − 4 (1 − ξa2 ) ⋯ ξair ⎥⎦
ir ir ir ir ir ir (19)

⎡ ξ bit 0 0 ⋯ 0 ⎤
⎢ ⎥
⎢ 1 − ξ bit 2
ξ bit 0 ⋯ 0 ⎥
⎢ ⎥
A uit =⎢ − ξ bit (1 − ξ b2it ) 1 − ξ b2it ξ bit ⋯ 0 ⎥
⎢ ⎥
⎢ ⋮ ⋮ ⋮ ⋱ ⋮ ⎥
⎢⎣ ( − ξ b ) Nb it − 2 (1 − ξ b2 ) ( − ξ b ) Nb it − 3 (1 − ξ b2 ) ( − ξ b ) Nb it − 4 (1 − ξ b2 ) ⋯ ξ bit ⎥⎦
it it it it it it (20)

and b yir and buit for ( i, r = 1, … , m) and (t = 1, … , p)) are vectors of


⎡ bi 0 ⎤ ⎡ bi
dimensions Nair and Nbit respectively, defined as follows: ⎢ u1
0 ⋯
⎥ 0 ⋯ 0 ⎤
⎢ y1 ⎥
⎢ 0 bi ⋯ 0 ⎥ ⎢ 0 bi ⋯ 0 ⎥
Bui′ = ⎢ u2
⎥, Biy′ = ⎢ y2 ⎥
1 − ξa2ir ⎡⎣ 1 − ξair (ξair )2 ⋯ ( − ξair ) Nair − 1⎤⎦
T
b iyr = (21) ⎢ ⋮ ⋮ ⋱ ⋮ ⎥ ⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0 0 … buip ⎦ ⎣ 0 0 … b iym ⎦ (28)

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
4 I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎

parameterization that would retain the benefits of high-order ARX


models. Toward this, it is proposed to parameterize the ARX

models using Laguerre basis. The presented MIMO ARX-Laguerre
model could represent coupled and decoupled systems and ensure
the complexity reduction with a recursive and easy representation.
In addition, the MIMO ARX-Laguerre model provides a rich re-

presentation of the dynamics and also provides a compact (par-
simonious in parameters) representation of high-order ARX
models. This measure can significantly reduce the number of
parameters to be identified and, as a consequence, time of iden-
∏ tification experiments can be significantly reduced.

3. Online parametric estimation of the MIMO ARX-Laguerre


Fig. 1. Parallel structure of the MIMO ARX-Laguerre model. model
and
3.1. Recursive identification of Fourier coefficients
⎧ m
⎪ Pny
i
= ∑ Nair
⎪ The proposed method is based on the minimization of a reg-
r=1
⎨ p ularised square error given, of the ith output, by the following
⎪ i
⎪ Pnu = ∑ Nbit expression:
⎩ t=1 (29)
h ⎛ m Nair
∑ ( yis (k) − CiT X i (k) )2 + αi ⎜⎜ ∑ ∑ ( gn, air (h − 1) − gn, air (h) )
2
The matrices A , i
and Bui Byi
are of dimensions (Pni × Pni ), (Pni × p) and Ji =
(Pni × m), respectively, with
k=1 ⎝ r= 1 n= 1
p Nb it ⎞
Pni = Pny ∑ ∑ ( gn, bit (h − 1) − gn, bit (h) ) ⎟⎟
i i 2
+ Pnu (30) +
t = 1 n= 1 ⎠ (35)
Thus, the relation (25) of the MIMO ARX-Laguerre model can be
implemented and numerically evaluated using a parallel structure where
as in Fig. 1.
From relation (25), the coupled multivariable system for p in-  yis , (i = 1 … m) is the ith output system.
puts and m outputs can be formulated by the following recursive  gn, air (h − 1) and gn, air (h) are the Fourier coefficients associated
representation: with the ith output at time instant (h − 1) and h, respectively.

 gn, bit (h − 1) and gn, bit (h) are the Fourier coefficients associated
⎪ X (k ) = AX (k − 1) + B u u (k − 1) + B y y (k − 1)
⎨ with the ith input at time instant (h − 1) and h, respectively.
⎪ T
⎩ y (k ) = C X (k ) (31)  αi , (αi > 0) is a regularisation constant which regulates the
importance that we give either to the quadratic error between
where X (k ) is a Pn-dimensional state vector defined as follows: yis (k ) and CiT X i (k ) or to that between the Fourier coefficients.
X (k ) = ⎡⎣ (X 1 (k ))T (X 2 (k ))T … (X m (k ))T ⎤⎦
T
(32) In matrix form, expression (35) becomes:
A, B u , B y and C are matrices of dimensions (Pn × Pn ), (Pn × p) and Ji = Yis (h) − Yi (h)
2
+ αi Ci (h − 1) − Ci (h) 2
(Pn × m), respectively, with:
T
m
= ( Yis (h) − Yi (h) ) ( Yis (h) − Yi (h) )
Pn = ∑ Pni + (Ci (h − 1) − Ci (h))T αi (Ci (h − 1) − Ci (h)) (36)
i=1 (33)
with
where Pni is given by relation (30).
Yis (h) = ⎡⎣ yis (1), …, yis (h) ⎤⎦
T
These matrices are defined as follows: (37)
⎡ 1⎤ ⎡ 1⎤
⎛ … 0 ⎞ ⎢ By ⎥ ⎛ C1 0⎞

A1 0
⎟ ⎢ Bu ⎥ ⎜
0 …

⎢ 2⎥ ⎢ 2⎥ 0 C2 0⎟
A2 C=⎜

A=⎜ 0 … 0 ⎟ B u = ⎢ Bu ⎥ B y = ⎢ By ⎥ Yi (h) = ⎡⎣ yi (1), …, yi (h) ⎤⎦
T
⎜ ⋱ 0 ⎟ ⎢ ⋮ ⎥ ⎜⋮ 0⎟ (38)

⋮ ⋮
⎟ ⎢ ⋮ ⎥ ⎜
⋮ ⋱

⎝ ⎢ B m⎥ ⎢ m⎥
0 0 … Am ⎠ ⎣ u⎦ ⎢⎣ By ⎥⎦ ⎝0 0 … Cm ⎠
(34) The vector Yi(h) can be also written as follows:
The MIMO ARX-Laguerre model of p inputs and m outputs is Yi (h) = X i (h) Ci (39)
characterised by m × (p + m) poles and (m × npi )-dimensional
vectors Ci. To ensure the reduction of the parameter number of the such as X i (h ) is defined as:
MIMO ARX-Laguerre model, the Laguerre poles have to be opti- X i (h) = ⎡⎣ X i (1), …, X i (h) ⎤⎦
T
(40)
mized. Thus, we propose in the next section an online procedure
based on the offline method proposed by Mbarek et al. [15] in the we can rewrite as, for (h = 1, … , M ):
MIMO case and Bouzrara et al. [16] in the SISO case. The Fourier
⎧ X i (h), if h = 1
coefficients are identified by a recursive method. ⎪
X i (h) = ⎨
It is important to note that for multivariable systems, the ⎪ ⎡ i i ⎤T
⎩ ⎣ X (h − 1), X (h) ⎦ , if h > 1 (41)
identification of high-order ARX models would require longer
experiments to be performed. This being undesirable from a where X i (h) is given by relation (23).
practical viewpoint, there is a need for a parsimonious And Ci (h − 1) and Ci(h) are the Fourier coefficients associated

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 5

with the ith output at time instant (h − 1) and (h) given by: the outputs and inputs, respectively. In this case, we consider the
⎡ following cost functions defined by Tanguy et al. [18]:
Ci (h) = ⎣ g0, ai1 (h), …, g Na − 1, ai1 (h), …, g0, aim (h), …, g Na − 1, aim (h),
i1 im
⎧ 1

⎤T ⎪ Jair = ∑ ngn2, air , (r = 1, …, m)
g0, bi1 (h), …, g Nb − 1, bi1 (h), …, g0, bip (h), …, g Nb − 1, aip (h) ⎦⎥ ⎪ ∥ hair ∥2
i1 ip (42) ⎨ n= 0
, (i = 1, …, m)

⎪ 1
Replacing Yi(h) in relation (36) by its expression given by (39), the ⎪ Jbit = ∥ h ∥2 ∑ ngn2, bit , ((t = 1, …, p)
regularized square error Ji is written as: ⎩ bit n= 0 (47)
T with
Ji = ( Yis (h) − X i (h) Ci (h) ) ( Yis (h) − X i (h) Ci (h) )
+ (Ci (h − 1) − Ci (h))T αi (Ci (h − 1) − Ci (h)) (43) ⎧ ∞
⎪ ∥ hair ∥2 = ∑ ha2ir (j), (r = 1, …, m)
The optimal parameter vectors Ci(h) are obtained by minimizing ⎪ j=0
⎨ ∞ , (i = 1, …, m)
the regularized square error Ji. As the criterion (36) is quadratic in ⎪
Ci(h), therefore its minimisation with respect to Ci(h) leads to a
2
⎪ ∥ hbit ∥ = ∑ hb2it (j), (t = 1, …, p)
⎩ j=0 (48)
global minimum:
∂Ji Consequently, and in a similar manner to that given by Bouzrara
= − 2 ( X iT (h) Yis (h) − X iT (h) X i (h) Ci (h))
∂Ci (h) et al. [16], the cost functions Js (s = air , bit ) given by (47) can be
expressed in an explicit manner according to the Laguerre poles
− 2αi (Ci (h − 1) − Ci (h)) (44)
ξs, (s = air , bit ), as follows:
i (h),
Then, the estimated of the ith Fourier coefficients vector, C ⎧
⎪ (1 + M1, air ) ξa2 − 2M2, air ξ air + M1, air
(i = 1, … , m), is: Ja = ir
, (r = 1, … , m)
⎪ ir
⎪ (1 − ξa2ir )
i (h) =
C ( X iT (h) X i (h) + αi I )−1 × ⎡⎣ X iT (h) Yis (h) + αi C
i (h − 1) ⎤
⎦ (45)

⎪ (1 + M1, b it ) ξ b2 − 2M2, b it ξ b it + M1, b it
, (i = 1, … , m)
⎪ it
Jb it = , (t = 1, … , p)
⎪ (1 − ξ b2 )
where I is the identity matrix of dimension (Nair + Nbit ). ⎩ it (49)
From (25) and (45), the recursive identification algorithm of
with
Fourier coefficients can be formulated as follows:
⎧ 1

Algorithm 1. Recursive identification of Fourier coefficients for a ⎪ M1, s = ∑ jhs2 (j)
MIMO system of p inputs and m outputs. ⎪ ∥ hs2 j=1
⎨ ∞ , (s = air , bit )
⎪ 1
1. Assume we have M inputs–outputs couples. ⎪ M2, s = ∑ jhs (j − 1) hs (j)
2. Choose the truncating orders Nair and Nbit and assume that ⎩ ∥ hs2 j=1 (50)
the Laguerre poles ξair and ξbit for (i, r = 1, … , m) and
By replacing, hs (j ) , (s = air , bit ), given by (8) in the relation (48) and
(t = 1, … , p) are optimized. by using the orthonormality property of the Laguerre functions,
3. Calculate the matrices Ai, Bui and Byi for (i, r = 1, … , m) and we obtain:
(t = 1, … , p).

4. Initialise Ci,0 = 0Nair + Nb ,1, yi (1) = 0 and Xi (1) = 0Nair + Nb ,1 for
it it ∥ hs ∥2 = ∑ gn2, s , (s = air , bit )
(i, r = 1, … , m) and (t = 1, … , p). n= 0 (51)
5. For h = 1, … , M The pole optimization of the MIMO ARX-Laguerre model consists
(a) Calculate Xi(h) for (i = 1, … , m).
in minimizing the cost functions Js , (s = air , bit ) for (r , i = 1, … , m)
(b) Form the matrix X i (h) given by (41). and (t = 1, … , p). The cost functions depend on the coefficients hs
(c) Compute the estimates Fourier coefficients at time in- (j), (s = air , bit ), of the MIMO ARX model, so we have to estimate
stant h using (45).
the structure as well as the parameters of such a model before
i (h) X i (h) for (i = 1, … , m).
(d) Compute y (h) = C i solving the optimization of the cost functions Js:
h = h + 1, go to step (a).
⎧ ∞
⎪ T1, s = ∑ (2n + 1) gn2, s
⎪ n= 0
3.2. Online pole optimization of the MIMO ARX-Laguerre model ⎨ ∞ (s = air , bit ) and n = 0, 1, …

⎪ T2, s = 2 ∑ ngn, s gn − 1, s
The MIMO ARX-Laguerre model is characterized by Pn para- ⎩ n= 0 (52)
meters, given by (33), depend on the truncating orders Nair and Nbit
The quantities T1, s and T2, s are related via their gradient with re-
of the development of the MIMO ARX model coefficients on La-
spect to the MIMO ARX-Laguerre poles ξs, (s = air , bit ) by the fol-
guerre bases. A parametric reduction is carried out if:
lowing relations [16]:
⎧ Na < na

⎨ ir ir
⎧ ∂T1, s 2
⎩ Nbit < nbit

(46) ⎪ = − T2, s
⎪ ∂ξs (1 − ξs2 )

The inequality given by (46) depends essentially on the choice of ⎪ 2, s
∂T 2
⎪ ∂ξ = − (1 T1, s
the MIMO ARX-Laguerre poles ξair and ξbit . Therefore, we propose ⎩ s − ξs2 ) (53)
an online optimization algorithm to optimize these poles and the
Fourier coefficients. By using the gradient of T1, s and T2, s given by (53) and the relation
We consider the separate optimization of the bases relating to (51), the quantities M1, s and M2, s , (s = air , bit associated to the
parameters hair (j ) and hbit (j ) of the MIMO ARX model associated to MIMO ARX-Laguerre model can be written as follows:

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
6 I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎

⎧ (1 + ξs2 ) T1, s + 2ξs T2, s 1 N (h) = ⎡ C


   ⎤T
⎪ M1, s = − C 0 ⎣ 1 (h) C2 (h) … Cm (h) ⎦ (63)
⎪ 2 (1 − ξs2 )∥ hs ∥2 2
⎨ The proposed method for the online parametric estimation of the
⎪ (1 + ξs2 ) T2, s + 2ξs T1, s
⎪ M2, s = MIMO ARX-Laguerre model consists in modifying the Laguerre
⎩ 2 (1 − ξs2 )∥ hs ∥2 (54) poles on the sliding interval [h + 2 − N0, h + 1] if the convergence
with criterion given by (62) is not satisfied. Once the poles are identi-
fied, the parameter adaptation technique is carried out on the
⎧ ∥ hs ∥2 ⎡
⎣ 2 (1 + ξs ) M1, s − 4ξs M2, s + (1 + ξs ) ⎤⎦
⎪ T1, s = 2 2 sliding window with the identified poles. This online parametric
⎪ (1 − ξs2 ) estimation algorithm is formulated as follows:

⎪ ∥ hs ∥2 ⎡ 2 ⎤
⎪ T2, s = 2 (1 − ξ 2 ) ⎣ −2ξs M1, s + (1 + ξs ) M2, s − ξs ⎦ Algorithm 2. Online parametric estimation of the MIMO ARX-
⎩ s (55) Laguerre model.
∂Ja
ir
The optimal poles ξopt , air and ξopt , bit solutions of ∂ξair
= 0 and 1. Assume we have M inputs/outputs data.
∂Jb
it
2. Fixed Nair , Nbit , αi, N0 and ε
= 0, respectively, are obtained from the Fourier coefficients of
∂ξb
it 3. On the interval [1, N0 ]
the MIMO ARX-Laguerre model, as: (a) Choose the initial values of the poles ξair , N0 and ξbit , N0
⎧ (b) Find the Fourier coefficients C i (N0 − 1), (i = 1, … , m),
⎪ ρair − ρa2 − 1 , if ρair > 1
ir
ξopt, air =⎨ using the Algorithm 1
⎪ρ + ρa2 − 1 , if ρair < − 1 4. For (N0 ≤ h ≤ M ) on the sliding interval [h + 2 − N0, h + 1]:
⎩ air ir (56)
(a) Estimate the Fourier coefficients C i (h), (i = 1, … , m),
using (59)
⎧ N (h) − C
(b) if ∥ C N (h − 1)∥2 ≤ ε then h = h + 1 and return
⎪ ρ bit − ρ b2 − 1, if ρ bit > 1 0 0
it
ξopt, bit =⎨ to step 4a
⎪ρ + ρ b2 − 1 , if ρ bit < − 1 Else
⎩ bit it (57)
i. Compute T1, air , T2, air , T1, bit and T2, bit
with ii. Compute ρair and ρbit
⎧ (1 + ξa2ir ) T1, air + 2ξair T2, air iii. Find the new poles ξair and ξbit , (i, r = 1, … , m) and
⎪ρ =
air (t = 1, … , p).
⎪ (1 + ξa2ir ) T2, air + 2ξair T1, air
⎨ iv. h = h + 1 and return to step 4a.
⎪ (1 + ξ b2it ) T1, bit + 2ξ bit T2, bit


ρ bit =
⎩ (1 + ξ b2it ) T2, bit + 2ξ bit T1, bit (58)

This method of MIMO ARX-Laguerre pole optimization is applied 4. Adaptive model based predictive control of MIMO systems
in the offline case by Mbarek et al. [15]. In this paper this meth- (AMPC MIMO)
od is extended to an online procedure, where we propose to
skip the observation sequences by sliding window at each This section considers the Adaptive Model based Predictive
increment of time instant. This window is of length Control of MIMO systems (AMPC MIMO) in which the parameters,
N0 ( << M )[h + 2 − N0, h + 1]. Then we combine the sliding win- poles and Fourier coefficients, are identified over a sliding window
dow with the recursive identification of the Fourier coefficients by applying Algorithm 2. The main advantage of this approach is
given by Algorithm 1 to propose an online parametric identifica- that the controller parameters are updated at each sample time,
tion of the MIMO ARX-Laguerre model. In this case the estimate of which usually means a faster response to operating mode of
Fourier coefficients i (h),
C (i = 1, … , m) is computed, for process changes. The controller update can also be monitored such
(h = N0, … , M ) , as:
that when there is a significant change in the process, a new
i (h) = ( X T (h) X i, N (h) + αi I )−1
C parameters model is identified then the controller is updated and
i, N0 0
implemented. In this paper a new combination of the online
× ⎡⎣ X iT, N0 (h) Yis, N0 (h) + αi C
i (h − 1) ⎤
⎦ (59) identification of the MIMO ARX-Laguerre model by the sliding
window procedure with the MPC MIMO is studied in an adaptive
with, for h = N0, … , M :
control context. The adaptive strategy is described with a block
Yis, N0 (h) = ⎡⎣ yis (h + 2 − N0 ), …, yis (h + 1) ⎤⎦
T diagram, as represented in Fig. 2.
(60)
The predictive control algorithm is based on the solving of
such as X i (h , N0 ) is defined as: optimization problem subject to a physical constraints on the in-
puts and the outputs due to the actuator technology, the control
X i (h) = ⎡⎣ X i (h + 2 − N0 ), …, X i (h + 1) ⎤⎦
T
(61) system security or the quality desired for the output of the con-
i , (i = 1, … , m), trolled process. In the case of MIMO system of m outputs,
The recursive identification of Fourier coefficients C
y1 (k ) , y2 (k ) , … , ym (k ), and p inputs, u1 (k ) , u2 (k ) , … , up (k ), for a
is evaluated by the quadratic error of the parameters Eqp in each
prediction horizon [k + 1, k + Np ] and a control horizon
sliding window as:
[k + 1, k + Nu ], (Nu ≤ Np ), the optimization problem is given as
N (h) − C
Eqp = ∥ C N (h − 1)∥2 ≤ ε (62) follows:
0 0

where C N (h) is a Pn-dimensional vector of MIMO ARX-Laguerre min (J (Δ U (k ))) (64)


0
Fourier coefficients.

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 7

Physical Disturbance The optimization problem (64) subject to (65) is a convex Quad-
constraints
ratic Programming problem (QP), for which the global solution is
Consigne Reference Optimization
unique [19]. As with all predictive control methods, where a
model -+ Algorithm Process
moving horizon strategy is applied, at instant kT only the first
optimized control u (k ) is applied to the system with:
f-step ahead
: Output prediction u (k ) = u (k − 1) + δ u (k ) (72)
: Control
: Predicted output
: Prediction error In this paper, we have considered adaptive predictive control
: Reference trajectory Online identiication of based on the flow chart in Fig. 3 and summarised by Algorithm 3.
MIMO ARX-Laguerre

Fig. 2. Bloc diagram of adaptive predictive control. Algorithm 3. AMPC MIMO using MIMO ARX-Laguerre model
algorithm.
⎧ ⎧ δu
⎪ ⎪ min ≤ δu1 (k + h) ≤ δ uM 1. Choose the m reference signal vectors wi, (i = 1, … , m), of
⎪⎨ ⋮ ; ∀ (h = 0, …, Nu − 1)
⎪ ⎪ δu dimension M.
≤ δup (k + h) ≤ δ uM
⎪ ⎩ min 2. Choose the prediction horizon Np, the control horizon Nu,
⎪⎧
⎪ u
⎪ ⎪ min
≤ u1 (k + h) ≤ uM the positive weights λ and the tracking error term γ.
subject to ⎨ ⎨ ⋮ ∀ (h = 0, 1, …, Nu − 1)
⎪⎪u 3. Choose the physical constraints given by (65).
⎩ ≤ up (k + h) ≤ uM
⎪ αi, ε and the width of the sliding window N0,
min
4. Fix Nair , Nbit ,
⎪⎧
⎪ ⎪ ymin ≤ y1 (k + h) ≤ yM N0 << M .
⎪⎨ ⋮ ∀ (h = 1, …, Np )
⎪⎪ 5. Assume that we have N0 inputs / outputs data system are

⎩ ⎩ ymin ≤ ym (k + h) ≤ yM (65) available.
6. Choose m × (m + p) stable poles ξair , N0 et ξbit , N0 .
where Δ U (k )) is the vector of the future increments control, de-
fined as: 7. On the interval [1, N0 ], estimate the Fourier coefficients
i (N0 − 1), (i = 1, … , m), using Algorithm 1.
C
Δ U (k ) = [δ u (k )T , δ u (k + 1)T , …, δ u (k + Nu − 1)T ]T (66) 8. Calculate the new poles ξair , N0 et ξbit , N0 from the relations
with, for j = 0, 1, … , Nu − 1, the vector δ u (k + j ) is defined as: (56) and (57).
9. Initialize the inputs u1 (h − 1) , … , up (h − 1) and fix h to N0.
δ u (k + j ) = [δu1 (k + j ), δu2 (k + j ), …, δut (k + j ), …, δup (k + j )]T (67)
10. While h < M
and, for t = 1, … , p, the tth increment control is defined as: (a) Using the optimal poles ξair , h and ξbit , h to compute the
matrices A , B u , B y and C defined by (34).
δut (k + j ) = ut (k + j ) − ut (k + j − 1) (68)
(b) Apply the inputs (u1 (h − 1) , … , up (h − 1)) to the system
In the predictive control, the optimization problem is based on the and read the outputs (y1 (h) , … , ym (h)).
minimization of prediction error between the predicted outputs (c) Compute the predicted outputs  y (k + f /k ) for
y^ , (i = 1, … , m) and the future setpoints w , (i = 1, … , m), then we
i i (f = 1, … , Np ) from relation (70).
can define the following quadratic criterion:
(d) Solve the optimization problem given by (64) for ob-
m Np tained (δu1 (h) , … , δup (h)).
∑ γi ∑ ⎡⎣ y^i (k + f /k) − wi (k + f ) ⎤⎦
2
J2 (ΔU (k )) = (e) Calculate the control signals (u1 (h) , … , up (h)) by (72).
i=1 h=1
p Nu− 1 (f) On the interval [h + 2 − N0, h + 1]:
+ ∑ λj ∑ [δuj (k + f )]2 i. Estimate the Fourier coefficients C i (h), (i = 1, … , m),
j=1 h=0 (69) using Algorithm 1.
where ii. if Eqp = ∥ CN (h) − C
N (h − 1)∥2 ≤ ε then (h = h + 1)
0 0
and go to step 10a.
 wi, (i = 1, …, m), are the reference trajectories, else, calculate the new poles ξair , h + 1 and ξbit , h + 1 from
 λj are positive weights generally considered constant and equal the relations (56) and (57), respectively, (h = h + 1) and go
to λ and γi are the weights on the tracking error term considered to step 10a.
constant and equal to γ.
 y^i (k + f /k ), (i = 1, …, m) are the f-step ahead prediction of
5. Simulation examples
outputs which is formulated by Mbarek et al. [15] from the
recursive representation of the MIMO ARX-Laguerre model
5.1. Numerical example
given by (31) as follows:
In order to validate the performances of the online identifica-
tion of the MIMO ARX-Laguerre presented in the paper, we con-
h
sider the following linear MIMO switched system with 2 inputs
y^ (k + f /k ) = y^ (k ) + C T ∑ (A + By C T )hδ X (k) and 2 outputs and composed of 3 submodels:
j=1
min (h, Nu ) ⎛ h−j ⎞ ⎧ x (k + 1) = Ai x (k ) + Bi u (k )

+ CT ∑ ⎜ ∑ (A + B y C T )l ⎟ B u δ u (k + j − 1) ⎨
⎜ ⎟ ⎩ y (k ) = Ci x (k ) + Di u (k )

(73)
j=1 ⎝ l= 0 ⎠ (70)

where δ X (k ) is the state vector increments, defined by: with u (k ) = [u1 (k ) , u2 (k )]T is the inputs vector, y (k ) = [y1 (k ) , y2 (k )]T
is the outputs, x(k) is the state vector at a sampling time k, i means
δ X (k ) = X (k ) − X (k − 1) (71) the mode of the system and the matrices Ai, Bi, Ci and Di are given by:

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
8 I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Fig. 3. Flow chart of AMPC MIMO using MIMO ARX-Laguerre model.

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 9

0.42 0.8
First output
Second output
0.4 0.6

0.38
0.4

0.36
0.2
0.34
0
0.32
−0.2
0.3

−0.4
0.28

0.26 −0.6
First input
Second input
0.24 −0.8
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
Number of iterations Number of iterations

Fig. 4. System inputs and outputs.

Fig. 5. Pole estimation of the MIMO ARX-Laguerre model.

⎧ α1 = α2 = 0.2.
⎛ 0.15 0.40 −0.65⎞ ⎛ −0.20 0.45⎞
⎪ ⎜ ⎟ ⎜ ⎟ These simulations are carried out with a sliding window
⎪ A1 = ⎜ −0.75 0.1 −0.35⎟ B1 = ⎜ −0.06 0 ⎟
⎪ N0 = 50 and a threshold ϵ = 0.02. To evaluate the performances of
⎨ ⎝ 0.20 0.70 0.20 ⎠ ⎝ 0.22 0 ⎠
⎪ the online identification Algorithm 2 of the MIMO ARX-Laguerre
⎪C = ⎜⎛ 0 0.40 0.45 ⎞ ⎛ 0 − 0.35⎞ model, we draw in Figs. 5–7 the evolution of the poles, the Fourier
⎪ 1 ⎟ D1 = ⎜ ⎟
⎩ ⎝ −1 −0.60 0.90⎠ ⎝ −1.70 −0.25⎠ (74) coefficients and the NMSE's, respectively. In Fig. 8 we plot the
evolution of the First and second outputs of the system and the
MIMO ARX-Laguerre model.
⎧ ⎛ 0.27 0.24 −0.55⎞ ⎛ −0.55 0 ⎞ We note from Figs. 5–8 the performance of the proposed online
⎪ ⎜ ⎟ ⎜ ⎟ identification procedure when the system is operating with sev-
⎪ A2 = ⎜ 0.24 0.65 0.30 ⎟ B2 = ⎜ −1.40 1 ⎟
⎪ eral modes. Indeed, we note the detection of the modes and the
⎨ ⎝ −0.55 0.30 0.27 ⎠ ⎝ 0.05 −0.72⎠
⎪ adaptation of the parameters models (Poles and Fourier coeffi-

⎪ C = ⎜ 0.70 1 −0.27 ⎞ ⎛ 2.15 0.25 ⎞
⎪ 2 ⎟ D2 = ⎜ ⎟ cients) where the outputs of the MIMO ARX-Laguerre model fol-
⎩ ⎝ −0.35 0 −1.10⎠ ⎝ 0 −0.36⎠ (75) lows faithfully the actual output of the system in spite of the
change of the operating modes of the system.

⎧ ⎛ 0.45 0.02 0.42 ⎞ ⎛ 0 0.15 ⎞


⎪ ⎜ ⎟ ⎜ ⎟ 5.2. Application of AMPC MIMO to the CSTR benchmark
⎪ A3 = ⎜ −0.17 0.53 0.20 ⎟ B3 = ⎜ 0.27 −0.46⎟

⎨ ⎝ 0.38 0.26 −0.66 ⎠ ⎝ 0.07 0.54 ⎠ 5.2.1. Process description

⎪ C = ⎛⎜ 0 0.60 0.28⎞

⎛ 0 −0.90⎞
D3 = ⎜ ⎟
The CSTR MIMO process used to validate the proposed AMPC


3
⎝ 0 0.86 0.45⎠ ⎝ 0 −0.85⎠ MIMO algorithm is described in [20] and illustrated in Fig. 9.
(76)
The process is modeled by the following equations:
To validate the online identification algorithm we use the input/
dCA q ⎛ −E ⎞
output signals given by Fig. 4 and we fixe a number of measures to = ( CA0 − CA ) − k 0 exp ⎜ ⎟
dt V ⎝ RT ⎠ (77)
M ¼1200, the truncating orders Nair = Nbit = 2, and the parameters

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
10 I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Fig. 6. Fourier coefficient variations of the MIMO ARX-Laguerre model.

Fig. 7. Square error variations of the first and second output.

(a) First output (b) Second output


Fig. 8. Validation of the online identification procedure.

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 11

Fig. 9. CSTR process.

Table 1
Steady state operating data.

Parameters Symbols Values

Product concentration CA 0.0882 mol/l


Reactor temperature T 441.2 K Fig. 11. CSTR inputs.
Coolant flow rate qc 100 l/min
Process flow rate q 100 l/min are the states (Fig. 10).
Feed concentration CA0 1 mol/l To evaluate the efficiency of Algorithm 3 of the MIMO ARX-
Feed temperature T0 350 K
Laguerre model, we propose to change the system dynamics at
Inlet coolant temperature TC0 350 K
CSTR volume V 100 l three operating points as given in Table 2.
Heat transfer term hA 7 × 105 cal/min K
Reaction rate constant K0 7.2 × 1010 /min 5.2.2. Online identification of CSTR process
Activation energy term E/R 1× 104 K The simulations are carried out with a sampling time T = 1 s.
Heat of reaction −ΔH 2 × 105 cal/mol The online identification procedure of the poles and the Fourier
Liquid densities ρ , ρC 1 × 103 g/l coefficient of the MIMO ARX-Laguerre model is applied for
Specific heat Cp, CPC 1 cla/g K Nair = Nbit = 2 (i, t , r = 1, 2). The windows length is N0 = 50. The
output is corrupted by an additive white and gaussian noise for a
signal to noise ratio equal to 30 dB. In the online identification we
will use the input signals given by Fig. 11.
To evaluate the performances of the online identification of the
MIMO ARX-Laguerre model, we plot in Figs. 12(a) and (b) the evo-
lution of the poles associated with the first output and the second
output, respectively. The evolution of the Fourier coefficient asso-
ciated with the first output and the second output is draw in Figs. 13
(a) and (b), respectively. In Figs. 14(a) and (b) we plot the evolution of
the validation of the first and second outputs of the MIMO ARX-La-
guerre model and the CSTR system, respectively. In Fig. 15 we draw
the evolution of the criterion J2. This evolution prove in the AMPC
Fig. 10. MIMO CSTR process.
MIMO ARX-Laguerre algorithm that the cost function value is limited
in time, i.e., J2 (k + 1) < J2 (k ) and the closed-loop system is stable.
Table 2 These results are obtained after some simulation tests by modifying
Operating points. the weighting factor γ1, γ2, λ1 and λ2 and the prediction and control
horizons Np and Nu. We note that the choice of the tuning parameters
Process flow rate Coolant flow rate Effluent concentration Temperature
Np, Nu, γ1, γ2, λ1 and λ2 is very important to guarantee the stability
q (l/min) qC (1/min) CA (mol/l) T (K)
performance of closed-loop system.
103 97 0.0748 445.3 We note from Figs. 14(a) and (b) the performance of the pro-
72 76 0.042 450.4 posed online identification algorithm in the identification of the
100 100 0.07 442.2 multiple modes systems. From Figs. 12(a) and (b) we observe the
detection of the operating modes of the system at 501 and 1001
and the adaptation of the poles and Fourier coefficients with the
dT q ⎛ −ΔH ⎞ ⎛ −E ⎞
= ( T0 − T ) + k 0 ⎜ ⎟ CA exp ⎜ ⎟ new operating modes.
dt V ⎝ ρCp ⎠ ⎝ RT ⎠
⎛ ρc Cpc ⎞ ⎡ ⎛ −hA ⎞ ⎤ 5.2.3. Adaptive predictive control of CSTR process
+⎜ ⎟ qc ⎢ 1 − exp ⎜ ⎟ ⎥ ( Tc 0 − T ) To validate the proposed MIMO adaptive predictive control
⎝ ρ Cp V ⎠ ⎣ ⎢ ⎝ qc ρc Cpc ⎠ ⎥⎦ (78) algorithm (MIMO AMPC) using the MIMO ARX-Laguerre model on
the adaptive control of the CSTR MIMO process we use the re-
The CSTR process has two state variables, namely the temperature ference signals given by Figs. 16(a) and (b).
T and the effluent concentration CA. The Jacketed CSTR is modeled The tuning parameters are:
using the parameters specified in Table 1 assuming the Feed flow
Np = 6, Nu = 1 and λ = γ = 0.5 (79)
rate and the Coolant Flow rate are the manipulated variables, the
effluent concentration and the Reactor temperature are the pro- The physical constraints on the control and the increment controls
cess outputs, effluent concentration and the Reactor temperature are chosen as follows:

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
12 I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Fig. 12. Online poles optimization.

Fig. 13. Online Fourier coefficients optimization.

Fig. 14. Evolution of the outputs of the MIMO ARX-Laguerre model and the CSTR system.

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 13

−0.4 ≤ δut (k + h) ≤ 0.4 ∀ (h = 0, …, Nu − 1) (80) In every step of the AMPC MIMO algorithm we optimize, over a
sliding window, the poles and the Fourier coefficients of the MIMO
ARX-Laguerre model using Algorithm 2. The variation of the op-
92 ≤ ut (k + h) ≤ 102 ∀ (h = 0, …, Nu − 1) (81) timal poles and the Fourier coefficients are given in Figs. 17 and 18,
respectively. In Figs. 19(a) and 22(b) we draw respectively the in-
for (t = 1, 2). crement controls and the controls signals. In Figs. 20(a) and (b) we
plot respectively the variation of the first reference signal and the
concentration output, and the second reference signal and the
temperature output.
The proposed AMPC MIMO ARX-Laguerre model is compared
with the adaptive model based predictive control using MIMO ARX
model (AMPC MIMO ARX). The development of the AMPC MIMO
ARX is presented in appendix Appendix A. To compare the per-
formances of the AMPC of the MIMO ARX-Laguerre model with
respect to the MIMO ARX one, we fix the same tuning parameters
given by relation (79), the same sliding window width N0 and the
same physical constraints given by relations (80) and (81). In
Fig. 21, we draw the variation of the outputs and the reference
signals obtained after applying the AMPC MIMO ARX Algorithm
with an order n ¼3. Fig. 22 gives the increment controls and the
controls signals. It can be mentioned the results given with the
MIMO ARX-Laguerre model are better than those given for MIMO
ARX model. This can be illustrated by both the convergence speed
of the outputs to the desired references and the computing time
Fig. 15. Evolution of the criterion J2. given by Figs. 23(a) and (b).

Fig. 16. Reference signals.

Fig. 17. Variation of the optimal poles.

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
14 I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Fig. 18. Variation of the Fourier coefficients.

Fig. 19. Variation of control signals in the case of AMPC MIMO ARX-Laguerre.

Fig. 20. Validation of the AMPC MIMO using ARX-Laguerre model.

6. Conclusion Predictive Control algorithm of the MIMO system (MIMO AMPC)


using the MIMO ARX-Laguerre model. This model is obtained by the
In this paper, we have proposed a new Adaptive Model based decomposition of the MIMO ARX model on the Laguerre

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 15

Fig. 21. Reference and output signals in the case of AMPC MIMO ARX.

Fig. 22. Variation of control signals in the case of AMPC MIMO ARX.

Fig. 23. Computing time.

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
16 I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎

orthonormal bases. The resulting model ensures a significant para- ⎡ 0 … 0 1 … 0 … 0 … 0 ⎤


⎢ ⎥
⎢ ⋮ ⋱ ⋮ ⋮ ⋱ ⋮ ⋮ 0 ⋱ 0 ⎥
meter number reduction with easy recursive representation. The two ⎢ 0 … 0 0 … 1 … 0 … 0 ⎥
⎢ ⎥
advantages of the MIMO ARX-Laguerre model allow us to develop an ⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
⎢ 0 … 0 0 … 0 … 1 … 0 ⎥
online identification algorithm to estimate the poles and the para- =⎢ ⋮ ⋱ ⋮ ⋮ ⋱ ⋮ ⋮ ⋮ ⋱ ⋮ ⎥ ∈ Rm . n × m . n
⎢ ⎥
meters by using a sliding window technique. We note that the ⎢ 0 … 0 0 … 0 … 0 … 1 ⎥
⎢ h a (n) … h a1m (n) h a11(n − 1) … h a1m (n − 1) … h a1m (1) ⎥
… h a11(1)
proposed online identification is very sensitive to the choice of the ⎢ 11 ⎥
⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
parameter αi which is inversely proportional to the SNR. By com- ⎢
⎣ h am1(n) … h amm (n) h am1(n − 1) … h amm (n − 1) … h am1(1) … h amm (1)⎦

(A.3)
bining the online identification algorithm with the predictive control
algorithm, an adaptive predictive control algorithm of MIMO system
is proposed in this paper. The two proposed algorithms, online ⎡ 0 … 0 … …0 0 ⎤
identification of MIMO ARX-Laguerre model and the adaptive pre- ⎢ ⋮ ⋱ ⋮ ⋮ ⋱⋮ ⋮ ⎥⎥
dictive control of MIMO system, are validated in a numerical simu- ⎢
⎢ 0 … 0 … …0 0 ⎥
lation and in a benchmark MIMO CSTR, of several operating points, ⎢ ⋮ ⋮ ⋮ ⋮ ⋮⋮ ⋮ ⎥
and the performances of the two algorithms are appreciated. ⎢ ⎥
⎢ 0 … 0 … …0 0 ⎥
m.n×p .n
0 = ⎢ ⋮ ⋱ ⋮ ⋮ ⋱⋮ ⋮ ⎥∈R
⎢ 0 … 0 … …0 0 ⎥
Appendix A. AMPC using ARX MIMO model ⎢ ⎥
⎢ hb11 (n) … hb1p (n) … hb11 (1)
… hb1p (1) ⎥
⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
A.1. Equations of predictions of ARX MIMO model ⎢ ⎥
⎢⎣ hbm1 (n) … hbmp (n) … hbm1 (1) … hbmp (1)⎥⎦ (A.4)
The ARX MIMO model given by (2) for (i, r = 1 ,.., m) can be
transformed to a classical state space form as follows:

⎛ y (k − n + 2) ⎞ ⎡ ⎤
⎜ 1 ⎟ ⎢ 0 … 0 1 … 0 … 0 … 0

⎜ ⋮ ⎟ ⎢ ⋮ ⋱ ⋮ ⋮ ⋱ ⋮ ⋮ 0 ⋱ 0 ⎥
⎜ y (k − n + 2)⎟ ⎢ ⎥
⎜ m
⎟ 0 … 0 0 … 1 … 0 … 0
⎜ ⋮ ⎟ ⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
⎜ y1 (k ) ⎟ ⎢ 0 0 0 0 1 0

⎜ ⎟= ⎢ … … … … ⎥
⎜ ⋮ ⎟ ⎢ ⋮ ⋱ ⋮ ⋮ ⋱ ⋮ ⋮ ⋮ ⋱ ⋮ ⎥
⎜ ⎟ ⎢ 0 … 0 0 … 0 … 0 … 1 ⎥
ym (k )
⎜ ⎟ ⎢ ⎥
⎜ y1 (k + 1) ⎟ ⎢ ha11 (n) … ha1m (n) ha11 (n − 1) … ha1m (n − 1) … ha11 (1) … ha1m (1) ⎥
⎜ ⎟ ⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
⎜⎜ ⋮ ⎟⎟ ⎢ ⎥
⎣ h a 1 (n) … hamm (n) ham1 (n − 1) … hamm (n − 1) … ham1 (1) … hamm (1) ⎦
⎝ ym (k + 1) ⎠ m

⎛ y (k − n + 1) ⎞ ⎡ ⎛ ⎞
⎜ 1 ⎟ ⎢ 0 … 0 … 0 … 0 ⎤ ⎜ u1 (k − n + 1) ⎟

⎜ ⋮ ⎟ ⎢ ⋮ ⋱ ⋮ ⋮ ⋮ ⋱ ⋮ ⎥⎜ ⋮ ⎟
⎜ y (k − n + 1)⎟ ⎢ 0 … 0 … 0 … 0 ⎥ ⎜ up (k − n + 1)⎟
⎜ m
⎟ ⎢ ⎜ ⎟
⎜ ⋮ ⎟ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥⎜ ⋮ ⎟
⎜ y (k − 1) ⎟ ⎢ 0 … 0 … 0 …

0 ⎥ ⎜ u1 (k − 1) ⎟
⎜ 1
⎟+⎢ ⋮ ⋮ ⎥⎜ ⎟
⎟ ⎢
⋱ ⋮ ⋮ ⋮ ⋱
⎜ ⋮ ⎜ ⋮ ⎟
⎜ y (k − 1) ⎟ ⎢ 0 … 0 … 0 … 0 ⎥ ⎜ u (k − 1) ⎟
⎟ ⎢ hb (n) ⎥⎜ p
⎜ m
… hb1p (n) … hb11 (1)
… hb1p (1) ⎥ ⎟
⎜ y1 (k ) ⎟ ⎢ 11 ⎜ u1 (k ) ⎟
⎜ ⎟ ⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥⎜ ⎟
⎟⎟ ⎢ h (n) ⎥⎜
⋮ ⋮
⎜⎜ … hbmp (n) … hbm1 (1) … hbmp (1)⎥⎦ ⎜ ⎟⎟
⎢ b
⎝ ym (k ) ⎠ ⎣ m1 ⎝ up (k ) ⎠ (A.1)

Then, we can write the flowing state space representation for the
⎛ y (k − n + 1) ⎞ ⎛ u1 (k − n + 1) ⎞
ARX MIMO model: ⎜ 1 ⎟ ⎜ ⎟
⎜ ⋮ ⎟ ⎜ ⋮ ⎟
⎜ y (k − n + 1)⎟ ⎜ up (k − n + 1)⎟
⎧  (k + 1) =   (k ) +  0  0 (k ) ⎜ m ⎟ ⎜ ⎟
⎜ ⎟

⎨ ⋮ ⎜ ⋮ ⎟
⎩ y (k ) =   (k )
⎪ ⎜ y (k − 1) ⎟ ⎜ u1 (k − 1) ⎟
(A.2)  (k ) = ⎜ 1 ⎟ ∈ R m . n × 1,  0 (k ) = ⎜ ⎟ ∈ Rp .n ×1
⎜ ⋮ ⎟ ⎜ ⋮ ⎟
⎜ y (k − 1) ⎟ ⎜ up (k − 1) ⎟
with: ⎜ m ⎟ ⎜ ⎟
⎜ y1 (k ) ⎟ ⎜ u1 (k ) ⎟
⎜ ⎟ ⎜ ⎟
⎜ ⋮ ⎟ ⋮
⎜⎜ ⎟⎟
⎜ ⎟ up (k )
⎝ ym (k ) ⎠ ⎝ ⎠ (A.5)

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 17

⎛0 … 0 … 0 … 0 1 0 … 0⎞ ⎛ y^ (k + 1) ⎞⎟ ⎛⎜ y^ (k )⎞⎟ ⎛   ⎞
⎜ ⎟ ⎜
0 … 0 … 0 … 0 0 1 … 0⎟ ⎜ ⎟
=⎜ ∈ Rm ×m . n ⎜ y^ (k + 2) ⎟ ⎜ y^ (k )⎟ ⎜  2 ⎟
⎜⋮ … ⋮ … ⋮ … ⋮ ⋮ ⋮ ⋱ ⋮⎟ ⎜ ⎟=⎜ ⎟+ δ  (k )
⎜ ⎟ ⎜ ⎟ ⎜ ⋮ ⎟ ⎜⎜⋮ ⎟
⎝0 … 0 … 0 … 0 0 0 … 1⎠ (A.6) ⋮ ⎟
⎜ ⎟ ⎜ ⎟ ⎝   p⎠
N
⎝ y^ (k + Np )⎠ ⎝ y^ (k )⎠
Considering the state space model given by (A.2) we can calculate ⎡   0 0m ×((N − 1) . p) ⎤
the prediction outputs equations of the MIMO ARX model [21]. ⎢ u

⎢   1 0m ×((Nu− 2) . p) ⎥
The state space model (A.2) can be written in incremental form as: ⎢ ⎥
⋮ ⋮
⎢ ⎥
⎧ ⎢   Nu− 2 0m ×(1. p) ⎥
⎪ δ  (k + 1) =  δ  (k ) +  0 δ  0 (k ) +⎢ ⎥ δ Np− 1(k + Np − 1)
⎨ ⎢
  Nu− 1

⎪ ^ ^
⎩ y (k ) = y (k − 1) +  δ  (k ) (A.7) ⎢   Nu ⎥
⎢ ⋮ ⎥
⎢ ⎥
where ⎣⎢   Np− 1 ⎦⎥ (A.13)

δ  (k + 1) =  (k + 1) −  (k ) or
δ 0 (k ) =  0 (k ) −  0 (k − 1) ^
Y (k ) = y^ (k ) + F Np δ  (k ) + G Np δ Np− 1(k + Np − 1) (A.14)
= ⎡⎣ δu1 (k − n + 1)…δup (k − n + 1)…δu1 (k )…δup (k ) ⎦ ⎤T (A.8)
with

From the first equation of relation (A.7), we can formulate the ⎛ y^ (k + 1) ⎞⎟


⎜ ⎛  ⎞
2-step prediction of the increment state vector: ⎜ ^ ⎟ ⎜ ⎟
^ y (k + 2)  2 ⎟
Y (k ) = ⎜ ⎟; F Np = ⎜ ;
⎜ ⋮ ⎟ ⎜ ⋮ ⎟
⎜ ⎟
δ  (k + 2) =  δ  (k + 1) +  0 δ 0 (k + 1) ⎜ ⎟ ⎝   p⎠N
⎝ y^ (k + Np )⎠
δ  (k + 2) = 2δ  (k ) +  0 δ 0 (k ) +  0 δ 0 (k + 1)
⎡   0 01 ×((N − 1) . p) ⎤
δ  (k + 2) = 2δ  (k ) + ⎣⎡ [ 0 0(m . n)× p] + [0(m . n)× p  0 ] ) ⎢ p

 ⎢   1 01 ×((Np− 2) . p) ⎥
1 ⎢ ⎥
G Np = ⎢ ⋮ ⋮ ⎥
⎡ δ T (k ) δu (k + 1)…δu (k + 1) ⎤T
⎣ 0 1 p ⎦ ⎢   Np − 2 01 ×(1. p) ⎥
 ⎢ ⎥
δ 1(k + 1) (A.9) ⎢⎣   Np− 1 ⎥⎦ (A.15)

by analogy, we can write δ  (k + 3) as: The output predictions vector Y^ (k ) can be decomposed into terms,
free components and forced components, as:
δ  (k + 3) = 3δ  (k ) ^ ^ ^
Y (k ) = Y fr (k ) + Y fo (k ) (A.16)
+ ⎡⎣ [ 1 0(m . n)× p] + [0(m . n)×(2p)  0 ] )
 with:
2
^
⎡ δ T (k + 1) δu (k + 2)…δu (k + 2) ⎤T Y fr (k ) = y^ (k ) + F Np δ  (k ) + Gfr δ fr (k ) (A.17)
⎣ 1 1 p ⎦

δ  2 (k + 2) (A.10)
^
Y fo (k ) = Gδ (k ) (A.18)
Then, the Np-step ahead prediction of  (k ) is written as:
where
δ  (k + Np ) =  Np δ  (k )
Gfr = G Np (: , 1: p. (n − 1));
+ ⎡⎣ [ N − 2 0(m . n)× p] + [0(m . n)×((Np− 1) p)  0 ] )
 G = G Np (: , p. (n − 1) + 1: p. Np ) (A.19)
N −1

⎡ δ T ⎤T ⎛ δ u (k − n + 1)⎞ ⎛ ⎞
⎣ Np− 2 (k + 1) δu1 (k + Np − 1)…δup (k + Np − 1) ⎦ δ u (k )
  ⎜ ⎟ ⎜ ⎟
δ fr (k ) = ⎜ ⋮ ⎟⎟; δ (k ) = ⎜ ⋮ ⎟
δ  Np − 1(k + Np − 1) (A.11) ⎜ ⎜ ⎟
⎝ δ u (k − 1) ⎠ ⎝ δ u (k + Np − 1)⎠ (A.20)
From relation (A.7) and by successive substitutions the f-step
ahead output predictions y^ (k + f /k ) is: A.2. Control optimization

f
The control action are obtained by minimizing the flowing
y^ (k + f /k ) = y^ (k ) +  ∑ δ  (k + j) quadratic criterion:
j=1 (A.12)
m Np

∑ qyi ∑ ⎡⎣ y^i (k + f /k) − wi (k + f ) ⎤⎦


2
For the prediction horizon [k + 1, k + Np ] and from the f-step J2 (u (k )) =
i=1 h=1
ahead output predictions given by (A.12) and from relations (A.9)– p Nu− 1
(A.11), we can also write the classical matrix notation of the output + ∑ quj ∑ [δuj (k + f )]2
predictions as: j=1 f =0 (A.21)

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i
18 I. Ben Abdelwahed et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎

where wi, i = 1, … , m, are the reference trajectories, qy(i) are the References
weights on the tracking error term and qu (j) are positive weights
generally considered constant and equal to qu. [1] Haitao Z, Zonghai C, Yongji W, Ming L, Ting Q. Adaptive predictive control
The quadratic criterion J2 (u (k )) given by (A.21) can be re- algorithm based on Laguerre functional model. Int J Adapt Control Signal
written in compact matrix form as: Process 2006;20(2):53–76.
[2] Ge S, Chenguang Y, Tong HL. Adaptive predictive control using neural network
^ ^ for a class of pure-feedback systems in discrete time. IEEE Trans Neural Netw
J2 (δ (k )) = ( Y (k ) − W (k ))T Q y ( Y (k ) − W (k )) 2008;19(9):1599–614.
+ δ (k )T Q u δ (k ) (A.22) [3] Yang W, Stephen B. Fast model predictive control using online optimization.
IEEE Trans Control Syst Technol 2010;18(2):267–78.
[4] Saghatoleslami N, Toroghi M. Controlling nonlinear processes, using Laguerre
where Y^ (k ) is the (m . Np )-dimensional vector of predicted outputs functions based adaptive model predictive control (AMPC) algorithm. J Chem
given by the relations (A.17) and the vector W(k) is defined as Petrol Eng 2011;45(1):47–55.
follows: [5] Li D, Xi Y. The synthesis of robust model predictive control with QP for-
mulation. Int J Modell Identif Control 2011;13(1/2):1–8.
W (k ) = [w1 (k + 1), …, wm (k + 1), …, w1 (k + Np ), …, wm (k + Np )]T (A.23) [6] Valencia Palomo G, Rossiter J. A novel PLC implementation of a predictive
controller based on Laguerre functions and multi-parametric solutions. IET
 (k ) is the vector of future control given by (A.20), and Qy and Qu Control Theory Appl 2012;6(8):1003–14.
[7] Qiang P, Tao Z, Qiumei C, Wang Y. Constrained model predictive control with
are positive-definite matrices of dimensions (m . Np ) and (p. Nu ),
economic optimization for integrating process. Can J Chem Eng 2015;93(8):
respectively, defined as follows: 1462–73.
[8] Bouzrara K, Mbarek A, Garna T. Non-linear predictive controller for uncertain
⎧ ⎧ ⎫
⎪ ⎪ ⎪ process modelled by GOBF-Volterra models. Int J Model Identif Control
⎪ Q = bloc _diag ⎪ ⎪
⎨ qy, …, qy ⎬
2013;19(4):307–22.
⎪ y ⎪  ⎪
[9] Rossiter J, Wang L. Exploiting Laguerre functions to improve the feasibility/
⎪ ⎪
⎩ Np times ⎪
performance compromise in MPC. In: Proceedings of the IEEE conference on
⎪ ⎭
⎨ decision and control; 2008. p. 4737–42.
⎪ ⎧ ⎫ [10] Wang L, Freemanb CT, Chaia S, Rogers E. Predictive-repetitive control with
⎪ ⎪ ⎪ constraints: from design to implementation. J Process Control 2013;23:
⎪ Q u = bloc _diag ⎨ qu, …, qu ⎬ 956–67.
⎪ ⎪  ⎪ [11] Douik A, Ghabi J, Messaoud H. Robust predictive control using a GOBF model

⎩ ⎩ N times ⎭ u (A.24) for MISO systems. Int J Comput Commun Control 2007;2(4):355–66.
[12] Badwe AS, Patwardhan SC, Gudi RD. Closed-loop identification using direct
where approach and high order ARX/GOBF-ARX models. J Process Control
2011;21:1056–71.
⎡q ⋯ 0 ⎤ ⎡q ⋯ 0 ⎤ [13] Ward MacArthur J. A new approach for nonlinear process identification using
⎢ y1 ⎥ ⎢ u1 ⎥ orthonormal bases and ordinal splines. J Process Control 2012;22:375–89.
qy = ⎢ ⋮ ⋱ ⋮ ⎥, qu = ⎢ ⋮ ⋱ ⋮ ⎥ [14] Tufa LD, Ramasamy M, Shuhaimi M. Improved method for development of
⎢ 0 ⋯ q ⎥ ⎢⎣ 0 ⋯ qup ⎥⎦ parsimonious orthonormal basis filter models. J Process Control 2011;21:
⎣ ym ⎦ (A.25)
36–45.
[15] Mbarek A, Bouzrara K, Garna T, Ragot J, Messaoud H. Laguerre-based model-
By replacing in relation (A.22) the predicted outputs Y^ (k ) by its ling and predictive control of multi-input multi-output systems applied to a
expression given by (A.16), the quadratic criterion J2 (δ  (k )) can be communicating two-tank system (CTTS). Trans Inst Meas Control 2015:1–14.
rewritten as: http://dx.doi.org/10.1177/0142331215613150.
[16] Bouzrara K, Garna T, Ragot J, Messaoud H. Online identification of the ARX
J2 (δ (k )) = δ (k )T ⎡⎣ G T Q y G + Q u ⎤⎦ δ (k ) model expansion on Laguerre orthonormal bases with filtres on model input
and output. Int J Control 2013;86(03):369–85.
^
+ 2 ( Y fr (k ) − W (k ))T Q y Gδ (k ) [17] Bouzrara K, Garna T, Ragot J, Messaoud H. Decomposition of an ARX model on
orthonormal Laguerre bases. ISA Trans 2012;51:848–60.
^ ^
+ ( Y fr (k ) − W (k ))T Q y ( Y fr (k ) − W (k )) (A.26) [18] Tanguy N, Morvan R, Vilb P, Calvez LC. Online optimization of the time scale in
adaptive Laguerre-based filters. IEEE Trans Signal Process 2000;48(4):1184–7.
Taking into account the constraints (65) on the input and [19] Fletcher R. Practical methods of optimization, vol. 2. John Wiley & Sons;
Chichester, 1981.
output signals leads to solve the optimization problem as follows: [20] Kamala N, Thyagarajan T, Renganathan S. Multivariable control of nonlinear
process using soft computing techniques. J Adv Inf Technol 2012;3(1):48–56.
min
Δ (k)∈ Ψ
( J2 (Δ (k)) ) (A.27) [21] Belda K, Böhm J. Adaptive generalized predictive control for mechatronic
systems. WSEAS Trans Syst 2006;5(8):1830–7.
where Ψ is the future controls admissible set formulated from
(65), [15], and J2 (Δ (k )) is the quadratic criterion given by (A.26).

Please cite this article as: Ben Abdelwahed I, et al. Adaptive MPC based on MIMO ARX-Laguerre model. ISA Transactions (2016), http:
//dx.doi.org/10.1016/j.isatra.2016.11.017i

You might also like