Você está na página 1de 15

Engineering Applications of Artificial Intelligence 24 (2011) 968–982

Contents lists available at ScienceDirect

Engineering Applications of Artificial Intelligence


journal homepage: www.elsevier.com/locate/engappai

Online set-point optimisation cooperating with predictive control of a yeast


fermentation process: A neural network approach
Maciej Ławryńczuk 
Institute of Control and Computation Engineering, Faculty of Electronics and Information Technology, Warsaw University of Technology, ul. Nowowiejska 15/19,
00-665 Warsaw, Poland

a r t i c l e i n f o a b s t r a c t

Article history: Online set-point optimisation which cooperates with model predictive control (MPC) and its applica-
Received 2 July 2010 tion to a yeast fermentation process are described. A computationally efficient multilayer control
Received in revised form system structure with adaptive steady-state target optimisation (ASSTO) and a suboptimal MPC
6 April 2011
algorithm are presented in which two neural models of the process are used. For set-point optimisation,
Accepted 11 April 2011
a steady-state neural model is linearised online and the set-point is calculated from a linear
Available online 10 May 2011
programming problem. For MPC, a dynamic neural model is linearised online and the control policy
Keywords: is calculated from a quadratic programming problem. In consequence of linearisation of neural models,
Process control the necessity of online nonlinear optimisation is eliminated. Results obtained in the proposed structure
Set-point optimisation
are comparable with those achieved in a computationally demanding structure with nonlinear
Model predictive control
optimisation used for set-point optimisation and MPC.
Yeast fermentation reactors
Neural networks & 2011 Elsevier Ltd. All rights reserved.
Linearisation

1. Introduction important in practice, they usually determine quality, economic


efficiency and safety. Moreover, MPC techniques are very efficient
Model predictive control (MPC) refers to a computer control when applied to process with many inputs and many outputs.
strategy in which an explicit dynamic model of the process is To maximise economic profits MPC algorithms cooperate with
used to predict its future behaviour over some time horizon set-point optimisation, the purpose of which is to calculate online
(Camacho and Bordons, 1999; Maciejowski, 2002; Rossiter, 2003; optimal set-points for MPC (Blevins et al., 2005; Brdys and
Tatjewski, 2007). Various MPC algorithms have been successfully Tatjewski, 2005; Engell, 2007; Tatjewski, 2007, 2008). Usually,
used for years in numerous advanced industrial applications (Qin the multilayer (hierarchical) structure is used in which the
and Badgwell, 2003), in particular in chemical engineering. It is control layer keeps process at given operating points and the
because MPC algorithms have a unique ability to take into optimisation layer calculates these set-points (Findeisen et al.,
account constraints imposed on both process inputs (manipulated 1980). It is also possible to integrate set-point optimisation and
variables) and outputs (controlled variables). Constraints are very MPC optimisation into one optimisation problem (Tvrzska de
Gouvea and Odloak, 1998; Tatjewski, 2007; Zanin et al., 2000,
2002). Alternatively, an integrated predictive optimiser and con-
straint supervisor can be used (Tatjewski et al., 2009). It provides
Abbreviation: ASSTO, adaptive steady-state target optimisation; LSSO, local
steady-state optimisation; LSSOþ MPC-NO, the ideal classical multilayer struc-
the control layer with set-points calculated for both optimality
ture with nonlinear set-point optimisation repeated at each sampling instant and and constraint handling.
the MPC algorithm with nonlinear optimisation; LSSO100 þMPC-NO, the realistic When the classical multilayer control structure is used, it is
classical structure with nonlinear set-point optimisation repeated 100 times less usually assumed that disturbances are slowly varying when com-
frequently than the MPC algorithm with nonlinear optimisation;
pared to the dynamics of the process (Tatjewski, 2007). In such a
LSSO100 þ ASSTOþ MPC-NPL, the structure with the ASSTO layer and the compu-
tationally efficient MPC-NPL algorithm, the LSSO layer is executed 100 less case, the steady-state nonlinear set-point optimisation problem can
frequently than the MPC-NPL algorithm; MFLOPS, millions of floating point be solved reasonably less frequently than the MPC controller
operations; MPC, model predictive control; MPC-NO, model predictive control executes. Provided that the dynamics of disturbances is much
with nonlinear optimisation; MPC-NPL, model predictive control with nonlinear slower than the dynamics of the plant, such an approach gives
prediction and linearisation; SSE, sum of squared errors; ss, steady-state; SSTO,
steady-state target optimisation.
good or satisfactory results. In many practical cases, however,
 Tel.: þ 48 22 234 76 73; fax: þ48 22 825 37 19. dynamics of disturbances is comparable with the process dynamics.
E-mail address: M.Lawrynczuk@ia.pw.edu.pl Very often disturbances, for example flow rates, properties of feed

0952-1976/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engappai.2011.04.007
M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982 969

and energy streams, etc., vary significantly and not much slower This paper is organised as follows. Section 2 shortly reminds the
than the dynamics of the controlled process. In such cases operation classical multilayer control structure, both set-point and MPC opti-
in the classical structure with frequency of set-point optimisation misation tasks are defined. Next, in Section 3, the multilayer structure
much lower than that of MPC may result in significant loss of with the ASSTO layer which cooperates with the suboptimal MPC
economic effectiveness (Tatjewski, 2007, 2008). Ideally, nonlinear algorithm is presented, linearisation of dynamic and steady-state
set-point optimisation should be repeated online as often as MPC is neural models is discussed. In Section 4 development of neural
activated. Because of high computational complexity, it is usually models of the yeast fermentation biochemical reactor is thoroughly
not possible. Moreover, nonlinear optimisation may terminate in described. The proposed structure is compared in terms of accuracy
local minima. Hence, nonlinear set-point optimisation is rarely used and computational burden with the structure in which nonlinear
online. optimisation is used online. Finally, Section 5 concludes the paper.
In order to reduce the computational complexity, in the simplest
case for set-point optimisation a constant linear steady-state model
derived from the dynamic model used in MPC can be used 2. The classical multilayer control structure
(Kassmann et al., 2000; Qin and Badgwell, 2003; Tatjewski, 2007,
2008). As a result, one obtains the steady-state target optimisation The standard multilayer system structure is depicted in Fig. 1
(SSTO) structure in which set-points are calculated from an easy to (Findeisen et al., 1980; Tatjewski, 2007, 2008). The basic control
solve linear programming problem. It can be solved online as layer is responsible for safe operation of the process. It has direct
frequently as MPC is activated. Naturally, for nonlinear processes access to input (manipulated) variables of the process. The
such an approach may give economically wrong operating points. supervisory control layer (also named the advanced control layer)
There are two solutions to this problem. It is possible to estimate and calculates online set-points for the basic control layer. The local
take into account the uncertainty in steady-state gain in the frame- steady-state optimisation (LSSO) layer calculates online econom-
work of a robust steady-state target calculation (Kassmann et al., ically optimal set-points for the supervisory control layer in such
2000). Alternatively, in the adaptive steady-state target optimisation a way that the production profit is maximised and constraints are
(ASSTO) structure the comprehensive nonlinear steady-state model satisfied. Each layer has a different frequency of intervention, the
is linearised online and set-point optimisation becomes a linear basic feedback control layer is the fastest. Plant-wide optimisa-
programming task (Ławryńczuk et al., 2008a; Qin and Badgwell, tion of complex processes is not considered in this paper, the
2003; Tatjewski, 2007, 2008). A yet another approach is to use reader is referred to Skogestad (2000, 2004).
piecewise-linear steady-state target optimisation (Ławryńczuk et al.,
2008b; Tatjewski, 2008; Tatjewski et al., 2006). The set-point 2.1. Set-point optimisation
optimisation layer can be also replaced by a neural network which
approximates the solution to the set-point optimisation problem The classical set-point optimisation problem is (for simplicity
(Ławryńczuk and Tatjewski, 2010). of presentation a single-input single-output process is assumed)
This paper details implementation of a computationally effi- (Tatjewski, 2007, 2008)
cient neural ASSTO structure which cooperates with a suboptimal min fJE ¼ cu uss cy yss g
ss
u
MPC algorithm. Two neural models of the process are used online.
For set-point optimisation, a steady-state neural model is line- subject to
arised online and the set-point is calculated from a linear pro- umin r uss r umax
gramming problem. For MPC, a dynamic neural model is linearised ymin ryss r ymax
online and the control policy is calculated from a quadratic yss ¼ f ss ðuss ,hss Þ ð1Þ
programming problem. In consequence of linearisation of neural
models, the necessity of online nonlinear optimisation is elimi-
nated. The proposed structure is applied to a simulated yeast
fermentation process. Fermentation is one of the most important
biochemical processes. Because properties of the process are non-
linear, the classical PID controller and the MPC algorithm based on
a linear model are unable to control the process efficiently as
demonstrated in Ławryńczuk (2008) and Nagy (2007). Results
obtained in the ASSTO structure are comparable with those
achieved in a computationally demanding structure with nonlinear
optimisation used for set-point optimisation and MPC.
Neural networks (Haykin, 1999; Ripley, 1996), due to their
advantages, can be successfully used for modelling and control of
nonlinear processes, e.g. Hussain (1999) and Nørgaard et al.
(2000). Neural networks are universal approximators (Hornik
et al., 1989), hence neural models are usually very precise. More-
over, unlike fundamental models (Luyben, 1990; Marlin, 1995)
(which are composed of algebraic and differential equations),
neural models have a simple structure and relatively a limited
number of parameters. Neural models directly describe input–
output relations of process variables, complicated systems of
equations do not have to be solved online in set-point optimisa-
tion and MPC. The literature concerned with MPC based on various
types of neural models is very rich, e.g. Alexandridis and Sarimveis
(2005), Hussain (1999), Ławryńczuk (2009), da Cruz Meleiro et al.
(2009), Nørgaard et al. (2000), Peng et al. (2007), Tatjewski (2007),
Yu and Gomm (2003) and references therein. Fig. 1. The classical multilayer control system structure.
970 M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982

where u is the input of the process (the manipulated variable), y is 3. The multilayer structure with adaptive steady-state target
the output (the controlled variable) and h is the measured (or optimisation and efficient MPC based on neural models
estimated) disturbance, the superscript ‘ss’ refers to the steady-
state. The function f ss : R2 -R denotes a steady-state model, cu, cy 3.1. The structure
represent prices resulting from economic considerations, umin,
umax, ymin, ymax denote constraints imposed on input and output The multilayer neural adaptive steady-state target optimisation
variables, respectively. (ASSTO) structure (Tatjewski, 2007, 2008) which cooperates with
MPC is depicted in Fig. 2. It consists of the nonlinear set-point
optimisation (LSSO) layer, the adaptive steady-state target optimi-
sation layer and the MPC layer. Two neural models are used online.
2.2. Model predictive control optimisation
A steady-state neural model is used in LSSO and in ASSTO layers,
whereas a dynamic model is used in MPC. Nonlinear LSSO set-
MPC algorithms are usually used in the supervisory control
point optimisation, in which the nonlinear optimisation problem
layer. It is because in MPC all necessary constraints can be easily
(1) is solved, is activated infrequently. At each sampling instant
taken into account. Moreover, in case of multivariable processes
ASSTO and MPC problems are solved. The ASSTO layer recalculates
with strong cross-coupling, a model-based MPC algorithm is a
the optimal set-point when the LSSO layer is not activated. Both
natural choice.
neural models are linearised online around the current operating
Let uss
lsso ðkÞ be the solution to the LSSO problem (1), carried out point. As a result, set-point optimisation (ASSTO) is a linear
for the current sampling instant k (or for some previous instant
programming task, MPC optimisation is a quadratic programming
and used for the current instant). Using the nonlinear steady-state
task. ASSTO and MPC layers have the same frequency of interven-
model, the value yss ss
lsso ðkÞ corresponding to ulsso ðkÞ is found. It is tion. In different words, the ASSTO layer, unlike the LSSO layer, is
next passed as the desired set-point to the predictive control
activated as frequently as the MPC algorithm is executed, i.e. at
optimisation problem (Camacho and Bordons, 1999; Maciejowski,
each sampling instant.
2002; Rossiter, 2003; Tatjewski, 2007) (for simplicity of presenta-
In the classical (i.e. nonadaptive) SSTO structure a constant
tion hard output constraints are used)
linear steady-state model is used for set-point calculation,
( ) whereas in the LSSO layer a comprehensive nonlinear steady-
X
N NX
u 1
2
min JMPC ðkÞ ¼ Jyss ^
lsso ðkÞyðk þpjkÞJ þ l JDuðk þ pjkÞJ2 state model is used. On the one hand it makes it possible to
DuðkÞ
p¼1 p¼0 formulate a linear programming SSTO optimisation problem, on
the other hand the linear model is usually only a rough estimate
subject to of the nonlinear one. In the presented structure the same steady-
state model is used in LSSO and ASSTO layers.
umin r uðk þ pjkÞ rumax , p ¼ 0, . . . ,Nu 1

Dumax r Duðk þ pjkÞ r Dumax , p ¼ 0, . . . ,Nu 1

ymin r yðkþ
^ pjkÞ r ymax , p ¼ 1, . . . ,N ð2Þ

In MPC at each sampling instant k (algorithm iteration) future


control increments are calculated

DuðkÞ ¼ ½DuðkjkÞ Duðk þ 1jkÞ . . . Duðk þ Nu 1jkÞT ð3Þ

where Duðkþ pjkÞ ¼ uðk þ pjkÞuðk þ p1jkÞ. It is assumed that


Duðk þ pjkÞ ¼ 0 for p Z Nu , where Nu is the control horizon. The
purpose of MPC is to minimise differences between the optimal
set-point yss ^
lsso ðkÞ and predictions yðk þ pjkÞ over the prediction
horizon N Z Nu , i.e. for p ¼1,y,N. Predictions are calculated from a
dynamic model of the process. Only the first element of the
determined sequence (3) is actually applied to the process

uðkÞ ¼ DuðkjkÞ þ uðk1Þ

At the next sampling instant, kþ1, output measurements are


updated and the whole procedure is repeated.
In the standard multilayer control structure two optimisation
problems are solved online: the LSSO task (1) and the MPC task
(2). When a comprehensive nonlinear steady-state model is used,
the first optimisation problem is nonlinear. Sometimes, for non-
linear processes linear MPC algorithms based on linear models
give quite good results. In such cases the MPC optimisation task is
in fact an easy to solve quadratic programming problem. Unfor-
tunately, for many processes, for example for the yeast fermenta-
tion reactor, linear MPC algorithms are inefficient (slow) as
demonstrated in Ławryńczuk (2008). If a nonlinear dynamic
model is used for control, the MPC task is also a nonlinear
optimisation problem. As a result, one obtains two nonlinear Fig. 2. The multilayer control system structure with adaptive steady-state target
optimisation tasks which must be solved online. optimisation (ASSTO) and MPC based on neural models.
M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982 971

3.2. Neural models where f : RnA þ nB þ nC tth þ 2 -R, integer numbers t, nB, th , nC, nA
(t r nB , th rnC ) define the order of dynamics. The function f is
Fig. 3 shows the structure of the steady-state neural model. A realised by the second feedforward neural network the structure
feedforward neural network with one hidden layer and a linear of which is shown in Fig. 4. Its output is
output (Haykin, 1999) is used. The output of the model is
X
K
K ss
X yðkÞ ¼ w20 þ w2i jðzi ðkÞÞ ð8Þ
ss ss ss
y ¼ f ðu ,h Þ ¼ ss
w2,ss
0 þ w2,ss
i
ðzss
j i Þ ð4Þ i¼1
i¼1
where K is the number of hidden nodes and zi(k) is the sum of
where the sum of inputs of the ith hidden node is inputs of the ith hidden node. From (7) one has
1,ss 1,ss ss 1,ss ss X
Iu Ih
X
zss
i ¼ wi,0 þ wi,1 u þwi,2 h ð5Þ
zi ðkÞ ¼ w1i,0 þ w1i,j uðkt þ1jÞ þ w1i,Iu þ j hðkth þ 1jÞ
j¼1 j¼1
Weights of the steady-state neural model are denoted by w1,ss
i,j
,
nA
X
ss
i¼1,y,K , j ¼0,1,2, and w2,ss
i , ss
i¼0,y,K , for the first and the þ w1i,Iu þ Ih þ j yðkjÞ ð9Þ
ss
second layer, respectively, K is the number of hidden nodes, j¼1
j: R-R is the nonlinear transfer function of hidden nodes
Weights of the dynamic neural model are denoted by w1i,j,
(e.g. hyperbolic tangent). From (4) and (5), the steady-state neural
i¼1,y,K, j ¼ 0, . . . ,nA þ nB þ nC tth þ 2, and w2i , i¼0,y,K, for
model is
the first and the second layer, respectively, Iu ¼ nB t þ1,
K ss
X Ih ¼ nC th þ1.
yss ¼ f ss ðuss ,hss Þ ¼ w2,ss
0 þ w2,ss
i
jðw1,ss
i,0
þ w1,ss
i,1
uss þ w1,ss
i,2
hss Þ ð6Þ
i¼1
3.3. Adaptive steady-state target optimisation based on neural
Fig. 4 shows the structure of the dynamic neural model. The models
model is described by the general discrete-time equation
yðkÞ ¼ f ðxðkÞÞ ¼ f ðuðktÞ, . . . ,uðknB Þ, If the steady-state model yss ¼ fss (uss,hss) is nonlinear, the set-
point optimisation problem (1) is also nonlinear. In order to
hðkth Þ, . . . ,hðknC Þ,
reduce computational complexity in the ASSTO layer the model
yðk1Þ, . . . ,yðknA ÞÞ ð7Þ is linearised online. Linearisation makes it possible to calculate
the set-point from a linear programming task.
The steady-state neural model (6) is linearised online around
the current state of the process determined by most recent
measurements u(k  1) and h(k). Using the Taylor series expan-
sion, the linear approximation of the nonlinear model is
yss ¼ f ss ðuss ,hss Þjuss ¼ uðk1Þ, hss ¼ hðkÞ þHðkÞðu
ss
uðk1ÞÞ ð10Þ

where

df ss ðuss ,hss Þ
HðkÞ ¼  ss ð11Þ
duss u ¼ uðk1Þ, hss ¼ hðkÞ

In contrast to the standard SSTO approach with a constant linear


model (Kassmann et al., 2000; Qin and Badgwell, 2003; Tatjewski,
2007, 2008), the discussed formulation uses a linearised model
derived online from the steady-state neural model. Taking into
Fig. 3. The steady-state neural model. account the structure of the steady-state neural model given
by (4) and (5), one has

XK ss
@jðzss 
i Þ
HðkÞ ¼ w2,ss  w1,ss ð12Þ
i¼1
i @zss
i
 ss ss
i,1
u ¼ uðk1Þ, h ¼ hðkÞ

If hyperbolic tangent is used as the nonlinear transfer function j


in the hidden layer of the model
@jðzss
i Þ 2
¼ 1tanh ðzss
i Þ ð13Þ
@zss
i

Taking into account (6), (10), (11), (12), the linearised steady-
state neural model is
K ss
X
yss ¼ w2,ss
0 þ w2,ss
i jðw1,ss 1,ss 1,ss
i,0 þ wi,1 uðk1Þ þwi,2 hðkÞÞ
i¼1

K ss
X @jðzss 
i Þ
þ w2,ss
i ss  w1,ss
i,1
ðuss uðk1ÞÞ ð14Þ
i¼1
@zi 
uss ¼ uðk1Þ, hss ¼ hðkÞ

Using the linearised steady-state model, from the nonlinear set-


point optimisation (LSSO) problem (1), one obtains the equivalent
linear programming ASSTO problem
min
ss
fJE ¼ cu uss cy yss g
Fig. 4. The dynamic neural model. u
972 M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982

subject to where l ¼1,y,nA, and

u min ss
ru ru max @f ðxðkÞÞ
bl ðkÞ ¼
@uðklÞ
ymin ryss r ymax 8
>
> 0, l ¼ 1, . . . , t1
<
yss ¼ f ss ðuss ,hss Þjuss ¼ uðk1Þ, ¼ X 2 @jðzi ðxðkÞÞÞ 1
ss K
hss ¼ hðkÞ þ HðkÞðu uðk1ÞÞ ð15Þ ð21Þ
>
> wi wi,lt þ 1 , l ¼ t, . . . ,nB
: i¼1
@zi ðxðkÞÞ
Let the solution to the ASSTO problem for the current sampling
instant be uss
assto ðkÞ. The linearised steady-state model (14) is next The linearisation point xðkÞ is determined by input, disturbance
used to calculate the optimal set-point yss assto ðkÞ. and output signals (measurements) corresponding to the argu-
ments of the nonlinear model (7). If hyperbolic tangent is used as
the nonlinear transfer function j in the hidden layer of the neural
3.4. Efficient nonlinear MPC-NPL algorithm based on neural models
dynamic model
If for prediction and optimisation of the control policy (3) a @jðzi ðxðkÞÞÞ 2
¼ 1tanh ðzi ðxðkÞÞÞ ð22Þ
^ þ pjkÞ are
nonlinear model is directly used in MPC, predictions yðk @zi ðxðkÞÞ
nonlinear functions of calculated increments DuðkÞ. In conse- Because the future output prediction (17) is a linear function
quence, the MPC optimisation problem (2) is a nonlinear task of the optimised control policy DuðkÞ, the MPC optimisation
which has to be solved online in real time. To eliminate the problem (2) becomes the following quadratic programming task:
necessity of nonlinear optimisation, the MPC algorithm with
nonlinear prediction and linearisation (MPC-NPL) (Ławryńczuk, min fJyref ðkÞGðkÞDuðkÞy0 ðkÞJ2 þ JDuðkÞJ2K
DuðkÞ, emin , emax
2007, 2009) is adopted (in this work the model and the algorithm
þ rmin Jemin J2 þ rmax Jemax J2 g
take into account the disturbance signal h). At each sampling
instant k a local linear approximation of the nonlinear neural subject to
dynamic model is calculated online. In consequence of linearisa- umin rJ DuðkÞ þ uðk1Þ r umax
tion, predictions depend in a linear way on the calculated control Dumax r DuðkÞ r Dumax
policy. Hence, the algorithm needs solving online a quadratic ymin emin rGðkÞ þ y0 ðkÞDuðkÞ rymax þ emax
programming problem, which can be easily done within a fore-
emin Z0, emax Z 0 ð23Þ
seeable time period.
Linearisation of the nonlinear dynamic model (7) derived at where
the current sampling instant k is
yref ðkÞ ¼ ½yss ss
lsso ðkÞ . . . ylsso ðkÞ
T
or yref ðkÞ ¼ ½yss ss
assto ðkÞ . . . yassto ðkÞ
T

yðkÞ ¼ bt ðkÞuðktÞ þ    þ bnB ðkÞuðknB Þ


ymin ¼ ½ymin . . . ymin T
a1 ðkÞyðklÞ    anA ðkÞyðknA Þ ð16Þ

where al(k) and bl(k) are time-varying coefficients of the line- ymax ¼ ½ymax . . . ymax T
arised model. It can be shown (Ławryńczuk, 2009) that if the are vectors of length N,
linear approximation (16) of the original nonlinear model (7) is
used for prediction in MPC, the output prediction vector yðkÞ^ ¼ umin ¼ ½umin . . . umin T
^ þ1jkÞ . . . yðkþ
½yðk ^ NjkÞT can be expressed as the sum of two parts
umax ¼ ½umax . . . umax T
^ 0
yðkÞ ¼ GðkÞDuðkÞ þ y ðkÞ ð17Þ
Dumax ¼ ½Dumax . . . Dumax T
The first part depends only on the future (on future increments
DuðkÞ which are calculated from the MPC optimisation problem uðk1Þ ¼ ½uðk1Þ . . . uðk1ÞT
(2) for the current sampling instant k), the second part is a
are vectors of length Nu, K ¼ l diagð1, . . . ,1Þ and
free trajectory vector y0 ðkÞ ¼ ½y0 ðkþ 1jkÞ . . . y0 ðk þ NjkÞT , which
2 3
depends only on the past. The dynamic matrix GðkÞ of dimension- 1 0 ... 0
ality N  Nu 61 1 ... 07
6 7
2 3 J ¼6 7
4^ ^ & ^5
s1 ðkÞ 0 ... 0
6 7 1 1 ... 1
6 s2 ðkÞ s1 ðkÞ ... 0 7
GðkÞ ¼ 6
6 ^
7
7 ð18Þ
4 ^ & ^ 5 are matrices of dimensionality Nu  Nu. If at the current sampling
sN ðkÞ sN1 ðkÞ . . . sNNu þ 1 ðkÞ instant the nonlinear LSSO layer is used, yref ðkÞ ¼ ½yss lsso ðkÞ . . .
T T
yss ref ss ss
lsso ðkÞ , if the ASSTO layer is used, y ðkÞ ¼ ½yassto ðkÞ . . . yassto ðkÞ .
is composed of step-response coefficients of the linearised model. To prevent infeasibility problems output constraints are softened
They are calculated recurrently (for j¼ 1,y,N) from coefficients of by means of slack variables (vectors emin and emax of length N),
the linearised model rmin and rmax 4 0 are weights (Maciejowski, 2002; Tatjewski,
2007).
minðj,n
X BÞ minðj1,n
X AÞ In the MPC-NPL algorithm for optimisation of the future
sj ðkÞ ¼ bi ðkÞ ai ðkÞsji ðkÞ ð19Þ
i¼1 i¼1
control policy a linearised model (16) is used (it makes it possible
to obtain a quadratic programming task (23)). On the other hand,
Taking into account the structure of the dynamic neural model for free trajectory calculation the full nonlinear neural model can
given by (8) and (9), coefficients of the linearised model (16) are be used. It is reflected in the name of the algorithm. From (8) the
calculated online from nonlinear prediction is

@f ðxðkÞÞ XK
@jðzi ðxðkÞÞÞ 1 X
K
al ðkÞ ¼  ¼ w2i wi,Iu þ Ih þ l ð20Þ ^ þpjkÞ ¼ w20 þ
yðk w2i jðzi ðk þ pjkÞÞ þ dðkÞ ð24Þ
@yðklÞ i¼1
@zi ðxðkÞÞ i¼1
M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982 973

where p¼1,y,N and IX


hf ðpÞ Ih
X
! þ w1i,Iu þ j hðkÞ þ w1i,Iu þ j hðkth þ 1j þ pÞ
X
K
j¼1 j ¼ Ihf ðpÞ þ 1
dðkÞ ¼ yðkÞyðkjk1Þ ¼ yðkÞ w20 þ w2i jðzi ðkÞÞ ð25Þ
IX
yp ðpÞ
i¼1
þ w1i,Iu þ Ih þ j y0 ðkj þpjkÞ
is the unmeasured disturbance estimate. Using (9) recurrently j¼1
one obtains nA
X
IX
þ w1i,Iu þ Ih þ j yðkj þ pÞ ð28Þ
uf ðpÞ
j ¼ Iyp ðpÞ þ 1
zi ðkþ pjkÞ ¼ w1i,0 þ w1i,j uðk t þ 1j þ pjkÞ
j¼1
Since future values of the disturbance are usually not known at
Iu
X the current sampling instant k, it is assumed that hðk þ pjkÞ ¼ hðkÞ
þ w1i,j uðkt þ 1j þ pÞ for p Z 1.
j ¼ Iuf ðpÞ þ 1

IX
hf ðpÞ

þ w1i,Iu þ j hðkth þ 1j þ pjkÞ 3.5. Summary of calculations


j¼1

Ih
Detailed configuration of the proposed structure is shown in
X
þ w1i,Iu þ j hðkth þ 1j þ pÞ Fig. 5. The ASSTO layer calculates the set-point as frequently as
j ¼ Ihf ðpÞ þ 1 the MPC-NPL algorithm is activated (at each sampling instant).
IX
yp ðpÞ
The LSSO layer is activated infrequently or for verification. Neural
þ w1i,Iu þ Ih þ j yðkj
^ þ pjkÞ models are linearised online: the ASSTO layer uses a linearised
j¼1 steady-state model (it solves a linear programming problem), the
nA
X MPC-NPL algorithm uses a linearised dynamic model (it solves a
þ w1i,Iu þ Ih þ j yðkj þpÞ ð26Þ quadratic programming problem).
j ¼ Iyp ðpÞ þ 1 At each sampling instant the following steps are repeated:
Auxiliary integer numbers are denoted by Iuf ðpÞ ¼ maxðmin
ðpt þ 1,Iu ÞÞ, Ihf ðpÞ ¼ maxðminðpth þ 1,Ih ÞÞ, Iyp ðpÞ ¼ minðp1,nA Þ. 1. Linearisation of the steady-state neural model: HðkÞ is found
Elements y0 ðkþ pjkÞ of the free trajectory vector y0 ðkÞ are from (12) and (13).
calculated recurrently online from the general nonlinear predic- 2. The optimal set-point ussassto ðkÞ is calculated from the ASSTO

tion equation (24) assuming only influence of the past optimisation problem (15). Next, the optimal set-point yss assto ðkÞ
is determined using the linearised model (14).
X
K
3. If the calculated set-point is verified, the set-point uss
y0 ðk þpjkÞ ¼ w20 þ w2i jðz0i ðk þpjkÞÞ þ dðkÞ ð27Þ lsso ðkÞ is

i¼1
also calculated from the LSSO nonlinear optimisation problem
(1). Next, the optimal set-point yss lsso ðkÞ is found using the
For free trajectory calculation no changes in the control signal nonlinear model (6).
from the current sampling instant k are assumed (i.e. 4. Linearisation of the dynamic neural model: coefficients al(k)
uðk þ pjkÞ: ¼ uðk1Þ for p Z0) and the nonlinear prediction is and bl(k) are obtained from (20) to (22). Next, step-response
replaced by the free trajectory. From (26) one obtains coefficients sj(k) which comprise the dynamic matrix GðkÞ (18)
IX
uf ðpÞ X
Iu are calculated from (19).
z0i ðk þ pjkÞ ¼ w1i,0 þ w1i,j uðk1Þ þ w1i,j uðkt þ 1j þ pÞ 5. Elements of the nonlinear free trajectory y0 ðkÞ are found using
j¼1 j ¼ Iuf ðpÞ þ 1 the dynamic neural model from (27), (28) and (25).

Fig. 5. Detailed configuration of the control structure with adaptive steady-state target optimisation and the MPC-NPL algorithm based on neural models.
974 M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982

6. The MPC-NPL quadratic programming problem (23) is solved dcX cS Fe


¼ mX CX expðKP cP Þ cX
to determine the optimal control policy DuðkÞ. As the set-point dt KS þ cS V
quantities yss ss
lsso ðkÞ or yassto ðkÞ calculated by LSSO or ASSTO
layers are used. dcP cS Fe
¼ mP CX expðKP1 cP Þ cP
7. The first element of the calculated vector is applied to the dt KS1 þ cS V
process, i.e. uðkÞ ¼ DuðkjkÞ þ uðk1Þ.
dcS 1 c
8. The iteration number is increased, i.e. k:¼kþ1, go to step 1. ¼ m cX S expðKP cP Þ
dt RSX X KS þcS
1 cS F Fe
 m cX expðKP1 cP Þ þ i cS,in  cS
RSP P KS1 þcS V V
4. Simulation results

4.1. Yeast fermentation reactor dcO2  Fe


¼ ðkla ÞðcO cO2 ÞrO2  cO2
dt 2
V
The considered yeast fermentation reactor is shown in Fig. 6. dTr F Fe rO2 DHr KT AT ðTr Tag Þ
The reactor is modelled as a continuous stirred tank with the ¼ i ðTin þ 273Þ ðTr þ 273Þ þ 
dt V V 32rr Cheat,r V rr Cheat,r
constant substrate feed flow, the outlet flow from the reactor
containing the product, the substrate and the biomass is also dTag Fag KT AT ðTr Tag Þ
constant. The reactor contains three distinct components: the ¼ ðT Tag Þ þ
dt Vj in,ag Vj rag Cheat,ag
biomass, which is a suspension of yeast fed into the system and
evacuated continuously, the substrate, which is the solution of The equilibrium concentration of oxygen in the liquid phase is
glucose feeding the micro-organism (Saccharomyces cerevisiae) P

cO ¼ ð14:60:3943Tr þ0:007714Tr2 0:0000646Tr3 Þ10 Hi Ii
and the product (ethanol), which is evacuated together with other 2

components. Together with yeast, inorganic salts are added. It is where the global effect of ionic strengths is
necessary for the formation of coenzymes. Inorganic salts have X m MNa mCaCO3 MCa
also a strong influence on the equilibrium concentration of Hi Ii ¼ 0:5HNa NaCl þ 2HCa
MNaCl V MCaCO3 V
oxygen in the liquid phase. !
The comprehensive fundamental model of the process is mMgCl2 MMg mNaCl mMgCl2 MCl
þ 2HMg þ0:5HCl þ2
thoroughly described in Nagy (2007), here the model is given in MMgCl2 V MNaCl MMgCl2 V
a compact form. State variables are: V—the volume of the mass of mCaCO3 MCO3
þ 2HCO3 þ 0:5HH 10pH þ0:5HOH 10ð14pHÞ
the reaction (l), cX—concentration of the biomass (yeast) (g/l), MCaCO3 V
cP—concentration of the product (ethanol) (g/l), cS—concentra-
tion of the substrate (glucose) (g/l), cO2 —concentration of oxygen
in the liquid phase (mg/l), Tr—temperature of the reactor (1C), The mass transfer coefficient for oxygen is
Tag—temperature of the cooling agent in the jacket (1C). The ðkla Þ ¼ ðkla Þ0 1:024Tr 20
reactor is described by the following continuous-time funda-
mental model containing seven nonlinear ordinary differential The rate of oxygen consumption is
equations: 1 cO2
rO2 ¼ mO2 cX
YO2 KO2 þ cO2
dV
¼ Fi Fe The maximum specific growth rate is
dt    
Ea1 Ea2
mX ¼ A1 exp  A2 exp 
RðTr þ273Þ RðTr þ273Þ
Parameters of the fundamental model are given in Table 1,
nominal operating conditions of the process are given in Table 2.
The fundamental model of the process is implemented in Matlab,
it is treated as the real process during simulations. Because model
equations are stiff (Ławryńczuk, 2008), a specialised solver for stiff
differential equations (ODE23S) is used (classical algorithms, e.g. the

Table 1
Parameters of the fundamental model.

A1 ¼ 9.5  108 (kla)0 ¼38 h  1 MMg ¼24 g/mol


A2 ¼ 2.55  1033 KO2 ¼ 8:86 mg=l MMgCl2 ¼ 95 g=mol
AT ¼ 1 m 2 KP ¼ 0.139 g/l MNa ¼23 g/mol
Cheat,ag ¼ 4.18 J g  1 K  1 KP1 ¼ 0:07 g=l MNaCl ¼ 58.5 g/mol
Cheat,r ¼ 4.18 J g  1 K  1 KS ¼ 1.03 g/l R ¼8.31 J mol  1 K  1
Ea1 ¼ 55 000 J=mol KS1 ¼ 1:68 g=l RSP ¼0.435
Ea2 ¼ 220 000 J=mol KT ¼ 3.6  105 J h  1 m  2 K  1 RSX ¼0.607
HCa ¼  0.303 mCaCO3 ¼ 100 g Vj ¼50 l
HCl ¼0.844 mMgCl2 ¼ 100 g YO2 ¼ 0:97 mg=mg
HCO3 ¼ 0:485 mNaCl ¼ 500 g DHr ¼ 518 kJ=mol O2
HH ¼  0.774 MCa ¼ 40 g/mol mO2 ¼ 0:5 h1
HMg ¼  0.314 MCaCO3 ¼ 90 g=mol mP ¼ 1:79 h1
HNa ¼  0.550 MCl ¼ 35:5 g=mol rag ¼ 1000 g=l
HOH ¼ 0.941 MCO3 ¼ 60 g=mol rr ¼ 1080 g=l
Fig. 6. The continuous yeast fermentation reactor.
M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982 975

Runge–Kutta 45 method, need small steps, and, in consequence, a classical PID controller and the MPC algorithm based on a linear
huge number of iterations). The fundamental steady-state model is model are unable to control the process efficiently (Ławryńczuk,
derived from the dynamic model by neglecting time and equating 2008; Nagy, 2007).
all time-derivatives to zero. Obtained steady-state model is com-
posed of a set of nonlinear equations. They are solved using the 4.2. Process modelling for set-point optimisation and MPC
Newton–Raphson method.
From the perspective of set-point optimisation and the MPC Since both dynamic and steady-state fundamental models are
algorithm, the fermentation process has one input (manipulated comprised of nonlinear equations which are solved iteratively, for
variable) Fag and one output (controlled variable) Tr. As discussed set-point optimisation and MPC simpler neural models are used.
in Nagy (2007), two main disturbances of the process can be Neural models can be easily used online because they are direct
considered: changes in the substrate concentration cS and in models, for a given input the output is calculated without the
temperature of the substrate flow entering the reactor Tin, but necessity of solving any equations.
only the second one has a significant effect on the process.
Both steady-state and dynamic properties of the reactor are 4.2.1. Steady-state modelling
nonlinear. The steady-state characteristics Tin(Fag,Tin) is shown in The fundamental steady-state model is used to generate
Fig. 7. In consequence of the nonlinear nature of the reactor, the random data from the whole domain of interest (0 l=h rFag r
100 l=h, 20 1C rTin r 30 1C). Next, different models are found.
Accuracy of models is defined by the sum of squared errors
Table 2 (SSE) performance function
Nominal operating conditions of the process.
Sss
X
2
cO2 ¼ 3:106953 mg=l Fi ¼51 l/h SSE ¼ ðyss ss
mod ðsÞy ðsÞÞ
cP ¼ 12.515241 g/l pH¼ 6 s¼1
cS ¼29.738924 g/l Tag ¼ 27.053939 1C
where yssmod ðsÞ denotes the output of the steady-state model (6),
cS,in ¼ 60 g/l Tin ¼ 25 1C
cX ¼0.904677 g/l Tin,ag ¼15 1C yss(s) denotes the steady-state target recorded during simulations
Fag ¼ 18 l/h Tr ¼29.573212 1C of the fundamental models, Sss is the number of steady-state
Fe ¼51 l/h V ¼1000 l samples. Since input, disturbance and output variables have a
different order of magnitude, they are scaled as
u ¼ 0:01ðFag Fag,nom Þ

h ¼ 0:2ðTin Tin,nom Þ

y ¼ 0:1ðTr Tr,nom Þ
where Fag,nom ¼ 18 l/h, Tin,nom ¼25 1C, Tr,nom ¼29.573212 1C corre-
spond to the nominal operating conditions of the process (Table 2).
In order to obtain and evaluate models, three sets of data are
used: training, validation and test sets (Ripley, 1996), each set has
2000 samples. The training data set is used only for neural network
training, the SSE performance index is minimised on this set. As a
result, parameters (weights) of neural networks are optimised. The
value of the SSE index for the validation data set is monitored
during training. To avoid overfitting training is terminated when
the validation error increases. Such an approach makes it possible
to obtain models which have good generalisation properties. Model
selection is performed taking into account only the value of SSE for
the validation data set. Finally, the third set (the test data set) is
Fig. 7. The steady-state characteristics of the reactor. used to assess generalisation abilities of the chosen model. It is

Table 3
Properties of linear, polynomial and neural steady-state models, for the chosen model SSEtest is calculated.

Model Number of parameters SSEtraining SSEvalidation SSEtest

3 3
Linear 3 4.8006  10 5.3011  10 –
Polynomial, 2nd order 9 1.0614  103 1.2505  103 –
Polynomial, 3rd order 16 2.7397  102 3.2342  102 –
Polynomial, 4th order 25 7.0009  101 8.3944  101 –
Polynomial, 5th order 36 1.7330  101 2.1443  101 –
Polynomial, 6th order 49 4.3895  100 5.6199  100 –
Polynomial, 7th order 64 1.1100  100 1.5700  100 –
Polynomial, 8th order 81 2.6659  10  1 4.9844  10  1 –
Polynomial, 9th order 100 6.0312  10  2 9.9902  10  2 –
Polynomial, 10th order 121 1.3253  10  2 6.7234  10  2 –
Neural, Kss ¼ 1 5 2.7364  101 3.0493  101 –
Neural, Kss ¼ 2 9 5.7613  10  1 6.5106  10  1 –
Neural, Kss ¼ 3 13 4.9481  10  2 6.6431  10  2 6.8547  10  2
Neural, Kss ¼ 4 17 6.8801  10  3 7.6947  10  3 –
Neural, Kss ¼ 5 21 5.2827  10  4 6.0908  10  4 –
976 M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982

important that the test data set is used neither for training nor for models in terms of accuracy and complexity. Due to nonlinearity
model selection. Therefore, the error on the test set gives an of the process, the linear model is very inaccurate. Polynomial
unbiased estimate of the generalisation error. models are more precise, but they need many parameters. Since
Neural models are trained using the BFGS (Broyden–Fletcher– neural networks are universal approximators (Hornik et al.,
Goldfarb–Shanno) optimisation algorithm (Bazaraa et al., 1993). 1989), they make it possible to approximate the steady-state
Weights of networks are initialised randomly and training is characteristics with very high accuracy. Finally, as a compromise
repeated 10 times for each model configuration (the multi-start between accuracy and complexity, the model with Kss ¼ 3 hidden
approach). Results presented next are the best obtained. nodes is chosen. Although it has only 13 parameters, its accuracy
Three classes of steady-state models are considered: linear, is comparable with that of the polynomial model of the 10th
polynomial and natural models. Table 3 compares obtained order which has as many as 121 parameters. Finally, the chosen

Fig. 8. Data used for dynamic modelling: (a) the training data set, (b) the validation data set and (c) the test data set.

Table 4
Properties of linear and neural dynamic models, for the chosen model SSEtest is calculated.

Model Number of parameters SSEtraining SSEvalidation SSEtest

Linear 6 6.1899  101 6.9683  101 –


Neural, K¼ 1 9 3.9827  101 3.9671  101 –
Neural, K¼ 2 17 7.8797  10  1 4.6090  10  1 –
Neural, K¼ 3 25 3.4225  10  1 2.3435  10  1 2.0981  10  1
Neural, K¼ 4 33 2.6516  10  1 3.2459  10  1 –
Neural, K¼ 5 41 2.8917  10  1 3.7258  10  1 –
M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982 977

Fig. 9. The output of the process (solid line with dots) vs. the output of the chosen neural dynamic model (dashed line with circles) and model errors: (a) the training data set,
(b) the validation data set and (c) the test data set.
978 M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982

neural model is tested. The SSEtest error is small, comparable with 4.3. Online set-point optimisation and MPC
the error for training and validation data sets. As the test set is
used neither for training nor for model selection, such a small Three systems structures are compared:
value means that the model generalises well.
1. LSSOþMPC-NO: the ideal classical multilayer structure with
nonlinear set-point optimisation (LSSO) repeated at each
4.2.2. Dynamic modelling sampling instant and the MPC-NO algorithm with full non-
To generate data for finding a dynamic model, the fundamental linear optimisation.
dynamic model is simulated open-loop in order to obtain training, 2. LSSO100 þASSTOþMPC-NPL: the proposed structure with the
validation and test data sets shown in Fig. 8. Each set contains ASSTO layer and the MPC-NPL algorithm, the LSSO layer is
4000 samples. The output signal contains small measurement executed 100 less frequently than the MPC-NPL algorithm.
noise. For dynamic models the SSE performance index is 3. LSSO100 þMPC-NO: the realistic classical structure with non-
linear set-point optimisation repeated 100 times less fre-
X
S quently than the MPC-NO algorithm.
SSE ¼ ðymod ðkjk1ÞyðkÞÞ2
k¼1
In the first case at each sampling instant two nonlinear optimisa-
where ymod ðkjk1Þ denotes the output of the dynamic tion problems are solved online at each sampling instant. For this
model for the sampling instant k calculated using signals up purpose sequential quadratic programming optimisation algo-
to the sampling instant k  1, y(k) are targets recorded during rithm is used (Bazaraa et al., 1993). In the second case a linear
simulations of the fundamental model, S is the number of programming set-point optimisation is solved in the ASSTO layer
samples. and a quadratic programming MPC-NPL task is solved. Both tasks
Models of different orders of dynamics have been considered. are solved at each sampling instant. The LSSO layer uses nonlinear
Finally, second-order neural models are used optimisation, but it is executed for verification very infrequently.
Finally, in the third structure the set-point is calculated by the
yðkÞ ¼ f ðuðk1Þ,uðk2Þ,hðk1Þ,hðk2Þ,yðk1Þ,yðk2ÞÞ ð29Þ LSSO layer by means of nonlinear optimisation (the ASSTO layer is
switched off). In order to illustrate usefulness of the ASSTO layer,
Table 4 shows accuracy and complexity of models with different
in this structure potentially the better MPC-NO algorithm rather
number of hidden nodes. Because data sets contain noise, when
than the suboptimal MPC-NPL one is used.
the model has too many weights (K 43), the SSE error for the
Three models of the process are used during simulation
validation data set increases. Hence, the model with K ¼3 hidden
experiments. The fundamental model is used as the simulated
nodes is chosen. The model has 25 weights. For the test data set
process, for set-point optimisation the steady-state neural model
the SSEtest error is also small, comparable with the error for
with Kss ¼3 hidden nodes is used, for MPC the dynamic neural
training and validation sets. Fig. 9 shows the output of the process
model with K ¼3 hidden nodes is used.
vs. the output of the neural model for training, validation and test
To maximise production, in set-point optimisation (LSSO and
data sets. Because the model is very precise, model errors are also
ASSTO layers) the minimised objective function is
depicted.
To show high accuracy of the neural model, a linear model is JE ¼ Fag
also found, it has the same arguments as the neural one (29).
Unfortunately, due to nonlinearity of the process, its accuracy is The same constraints imposed on the manipulated variable are
very low. For the linear model SSEvalidation ¼6.9683  101, whereas used in set-point optimisation and MPC
for the chosen neural models SSEvalidation ¼2.3435  10  1. Fig. 10 min
Fag ¼ 0 l=h, max
Fag ¼ 100 l=h
shows the output of the process vs. the output of the linear model
for training and validation data sets, no comparison for the test Additionally, in set-point optimisation and MPC a constraint is
data set is given since the linear model is not finally selected. In imposed on the controlled output to guarantee that Tr r Trmin ,
contrast to the neural model (Fig. 9), model errors are significant. where the value of Trmin can be adjusted by the operator of the

Fig. 10. The output of the process (solid line) vs. the output of the linear model (dashed line): (a) the training data set and (b) the validation data set.
M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982 979

process. The disturbance scenario is The disturbance changes sinusoidally from the sampling instant
( k¼100, its amplitude is 5. The disturbance signal is shown in
Tin,nom k o 100
Tin ðkÞ ¼ Fig. 11.
Tin,nom þ 5ðsinð2ðk100ÞÞÞ l=h 100 r k r600
Parameters of MPC (MPC-NPL and MPC-NO algorithms) are:
N¼10, Nu ¼2, l ¼ 2, the same as in Ławryńczuk (2008), rmin ¼
rmax ¼ 20. Fig. 12 depicts simulation results of three compared
systems structures for Tmin
r ¼25. At the beginning of simulations
the LSSO layer calculates the optimal set-point (Fmin
ag ¼65.3764 l/h,
Tmin
r ¼25 1C). Because the process is started using nominal
conditions (Table 2), the MPC algorithm needs some time (approxi-
mately 70 sampling instants) to steer the process to the desired
set-point. In the LSSO100 þASSTOþMPC-NPL structure the LSSO
layer is also activated for k¼100, 200, 300, 400, 500, 600, for all
remaining iterations the set-point is calculated using the ASSTO
layer. Fig. 13 shows simulation results for Tmin
r ¼24 1C, whereas
Fig. 14 for Tmin
r ¼26 1C.
Considering obtained simulation results, two observations can
be made:

1. Trajectories obtained in the proposed structure with the


ASSTO layer and the suboptimal MPC-NPL algorithm (the
LSSO100 þASSTOþMPC-NPL structure) are very close to those
obtained in the ideal but unrealistic case with nonlinear set-
point optimisation and nonlinear MPC optimisation repeated
Fig. 11. The disturbance scenario. at each sampling instant (the LSSOþMPC-NO structure).

Fig. 12. Simulation results for Tmin


r ¼25 1C: (a) the ideal LSSOþ MPC-NO structure (solid line) and the proposed LSSO100 þASSTO þMPC-NPL structure (dashed line) and
(b) the realistic LSSO100 þMPC-NO classical structure.
980 M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982

Fig. 13. Simulation results for Tmin


r ¼24 1C: (a) the ideal LSSOþ MPC-NO structure (solid line) and the proposed LSSO100 þ ASSTO þMPC-NPL structure (dashed line) and
(b) the realistic LSSO100 þMPC-NO classical structure.

2. The realistic classical structure with infrequent nonlinear set- is 659.89 MFLOPS, in the LSSO100 þ ASSTOþMPC-NPL structure it
point optimisation (LSSO100 þMPC-NO) gives significantly dif- is reduced to 61.32 MFLOPS.
ferent trajectories. It means that the ASSTO layer is really
necessary.
5. Conclusions

For the whole simulation horizon (600 sampling instants) the The described structure with adaptive steady-state target
economic performance index optimisation (ASSTO) and the MPC-NPL algorithm is computa-
tionally efficient and gives trajectories very similar to those
X
600 X
600
obtained when for set-point optimisation and MPC nonlinear
JE ¼ JE ðkÞ ¼ ðFag ðkÞÞ
k¼1 k¼1 optimisation is used online. For set-point optimisation and MPC
steady-state and dynamic neural models are linearised online. In
is calculated after simulations. Table 5 gives values of the index JE consequence, set-point optimisation is an easy to solve linear
obtained in the ideal multilayer LSSOþMPC-NO structure and in programming task, MPC optimisation is a quadratic programming
the proposed LSSO100 þASSTOþMPC-NPL structure for different problem. These optimisation tasks can be efficiently solved
constraint values Tminr . As Figs. 12–14 suggest, both structures online, the necessity of nonlinear optimisation is eliminated.
yield similar economic performance. For example, for Tmin r ¼25, in This paper recommends using neural models for set-point
the LSSOþMPC-NO structure JE ¼  2.2778  106, in the LSSO100 þ optimisation and MPC. Development of steady-state and dynamic
ASSTOþMPC-NPL structure JE ¼ 2.3137  106. neural models is detailed. Although a dynamic fundamental (first-
Although trajectories obtained in the ideal multilayer principle) model of the process is available, it is not suitable to be
LSSOþMPC-NO structure and in the proposed LSSO100 þASSTOþ directly used in MPC because it is composed of a set of stiff
MPC-NPL structure are vary similar, the difference in computa- differential equations which must be solved online. The neural
tional burden is very big. Table 6 shows computational complex- model is very precise and has a limited number of parameters.
ity of these two structures in MFLOPS (millions of floating point Moreover, its structure is simple, regular. In consequence,
operations) for different constraint values Tmin
r . For example, for it can be easily used online in the computationally efficient
Tmin
r ¼25, in the LSSOþMPC-NO structure the computational cost MPC-NPL algorithm, complicated systems of nonlinear differential
M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982 981

Fig. 14. Simulation results for Tmin


r ¼26 1C: (a) the ideal LSSOþ MPC-NO structure (solid line) and the proposed LSSO100 þASSTO þMPC-NPL structure (dashed line) and
(b) the realistic LSSO100 þMPC-NO classical structure.

Table 5 equations do not have to be solved online. Analogously, for set-


Values of the economic performance index JE obtained in the ideal multilayer point optimisation a neural approximate model is used,
LSSOþ MPC-NO structure and in the proposed LSSO100 þ ASSTOþ MPC  NPL
it is not necessary to solve online a set of nonlinear algebraic
structure.
equations comprising the comprehensive fundamental steady-
Tmin
r LSSO þMPC  NO structure LSSO100 þASSTO state model.
þMPC  NPL structure

22  3.4820  106  3.5504  106


23  3.0695  106  3.1187  106
24  2.6778  106  2.7209  106
Acknowledgement
25  2.2778  106  2.3137  106
26  1.9762  106  1.9925  106 The work presented in this paper was supported by Polish
27  1.6085  106  1.6073  106 national budget funds for science.
28  9.7405  105  9.7302  105

Table 6 References
Computational complexity in terms of floating point operations (MFLOPS) of the ideal
multilayer LSSOþMPC NO structure and of the proposed LSSO100 þ ASSTO
þ MPC NPL structure. Alexandridis, A., Sarimveis, H., 2005. Nonlinear adaptive model predictive control
based on self-correcting neural network models. AIChE J. 51, 2495–2506.
Bazaraa, M.S., Sherali, J., Shetty, K., 1993. Nonlinear Programming: Theory and
Tmin
r (1C) LSSO þMPC  NO structure LSSO100 þASSTO
Algorithms. John Wiley & Sons, Hoboken, New Jersey.
þMPC  NPL structure
Blevins, T.L., McMillan, G.K., Wojsznis, M.W., 2005. Advanced Control Unleashed.
ISA.
22 353.96 46.38 Brdys, M.A., Tatjewski, P., 2005. Iterative Algorithms for Multilayer Optimizing
23 407.41 49.75 Control. Imperial College Press/World Scientific, London.
24 479.28 52.37 Camacho, E.F., Bordons, C., 1999. Model Predictive Control. Springer, London.
25 659.89 61.32 da Cruz Meleiro, L.A., José, F., Zuben, V., Filho, R.M., 2009. Constructive learning
26 758.71 68.34 neural network applied to identification and control of a fuel–ethanol
27 1011.34 75.22 fermentation process. Eng. Appl. Artif. Intell. 22, 201–215.
28 927.71 85.39 Engell, S., 2007. Feedback control for optimal process operation. J. Process Control
17, 203–219.
982 M. Ławryńczuk / Engineering Applications of Artificial Intelligence 24 (2011) 968–982

Findeisen, W.M., Bailey, F.N., Brdyś, M., Malinowski, K., Tatjewski, P., Woźniak, A., Nørgaard, M., Ravn, O., Poulsen, N.K., Hansen, L.K., 2000. Neural Networks for
1980. Control and Coordination in Hierarchical Systems. John Wiley & Sons, Modelling and Control of Dynamic Systems. Springer, London.
Chichester. Peng, H., Yang, Z.J., Gui, W., Wu, M., Shioya, H., Nakano, K., 2007. Nonlinear system
Haykin, S., 1999. Neural Networks—A Comprehensive Foundation. Prentice-Hall, modeling and robust predictive control based on RBF-ARX model. Eng. Appl.
Englewood Cliffs. Artif. Intell. 20, 1–9.
Hornik, K., Stinchcombe, M., White, H., 1989. Multilayer feedforward networks are Qin, S.J., Badgwell, T.A., 2003. A survey of industrial model predictive control
universal approximators. Neural Networks 2, 359–366. technology. Control Eng. Pract. 11, 733–764.
Hussain, M.A., 1999. Review of the applications of neural networks in chemical Ripley, B.D., 1996. Pattern Recognition and Neural Networks. Cambridge Univer-
process control—simulation and online implementation. Artif. Intell. Eng. 13,
sity Press, Cambridge.
55–68.
Rossiter, J.A., 2003. Model-Based Predictive Control. CRC Press, Boca Raton.
Kassmann, D.E., Badgwell, T.A., Hawkins, R.B., 2000. Robust steady-state target
Skogestad, S., 2004. Control structure design for complex chemical plants. Comput.
calculation for model predictive control. AIChE J. 46, 1007–1024.
Chem. Eng. 28, 219–234.
Luyben, W.L., 1990. Process Modelling, Simulation and Control for Chemical
Skogestad, S., 2000. Plantwide control: the search for the self-optimizing control
Engineers. McGraw-Hill, New York.
Ławryńczuk, M., Tatjewski, P., 2010. Approximate neural economic set-point structure. J. Proc. Control 10, 487–507.
optimisation for control systems. In: Rutkowski, L., Scherer, R., Tadeusiewicz, Tatjewski, P., Ławryńczuk, M., Marusak, P., 2009. Integrated predictive optimiser
R., Zadeh, L.A., Zurada, J. (Eds.), Proceedings of the 10th International and constraint supervisor for processes with basic feedback control algorithm.
Conference on Artificial Intelligence and Soft Computing, ICAISC 2010, Zako- In: Proceedings of the European Control Conference, ECC 2009, Budapest,
pane, Poland, Lecture Notes in Artificial Intelligence, vol. 6114. Springer- Hungary, pp. 3359–3364.
Verlag, Berlin Heidelberg, pp. 305–312. Tatjewski, P., 2008. Advanced control and on-line process optimization in multi-
Ławryńczuk, M., 2009. Neural networks in model predictive control. In: Nguyen, layer structures. Annu. Rev. Control 32, 71–85.
N.T., Szczerbicki, E. (Eds.), Studies in Computational Intelligence, Intelligent Tatjewski, P., 2007. Advanced Control of Industrial Processes, Structures and
Systems for Knowledge Management, vol. 252. Springer, Berlin, pp. 31–63. Algorithms. Springer, London.
Ławryńczuk, M., 2008. Modelling and nonlinear predictive control of a yeast Tatjewski, P., Ławryńczuk, M., Marusak, P., 2006. Linking nonlinear steady-state
fermentation biochemical reactor using neural networks. Chem. Eng. J. 145, and target set-point optimisation for model predictive control. In: Proceedings
290–307. of the International Control Conference ICC 2006, Glasgow, Scotland, United
Ławryńczuk, M., Marusak, P., Tatjewski, P., 2008a. Cooperation of model predictive Kingdom, CD-ROM, paper 171.
control with steady-state economic optimisation. Control Cybernet. 37, Tvrzska de Gouvea, M., Odloak, D., 1998. One-layer real time optimization of LPG
133–158.
production in the FCC unit: procedure, advantages and disadvantages. Com-
Ławryńczuk, M., Marusak, P., Tatjewski, P., 2008. Piecewise linear steady-state
put. Chem. Eng. 22, S191–S198.
target optimization for control systems with MPC: a case study. In: Proceed-
Yu, D.L., Gomm, J.B., 2003. Implementation of neural network predictive control to
ings of the 17th IFAC World Congress, Seoul, South Korea, pp. 13169–13174.
Ławryńczuk, M., 2007. A family of model predictive control algorithms with a multivariable chemical reactor. Control Eng. Pract. 11, 1315–1323.
artificial neural networks. Int. J. Appl. Math. Comput. Sci. 17, 217–232. Zanin, A.C., Tvrzska de Gouvea, M., Odloak, D., 2002. Integrating real-time
Maciejowski, J.M., 2002. Predictive Control with Constraints. Prentice-Hall, optimization into model predictive controller of the FCC system. Comput.
Harlow. Chem. Eng. 26, 819–831.
Marlin, T.E., 1995. Process Control. McGraw-Hill, New York. Zanin, A.C., Tvrzska de Gouvea, M., Odloak, D., 2000. Industrial implementation of
Nagy, Z.K., 2007. Model based control of a yeast fermentation bioreactors using a real-time optimization strategy for maximizing production of LPG in a FCC
optimally designed artificial neural networks. Chem. Eng. J. 127, 95–109. unit. Comput. Chem. Eng. 24, 525–531.

Você também pode gostar