You are on page 1of 15

Available online at www.sciencedirect.

com

Computers and Chemical Engineering 32 (2008) 18771891

A Tabu search-based algorithm for mixed-integer nonlinear problems


and its application to integrated process and control system design
Oliver Exler 1 , Luis T. Antelo, Jose A. Egea, Antonio A. Alonso, Julio R. Banga
Process Engineering Group, IIM-CSIC, C/ Eduardo Cabello 6, 36208 Vigo, Spain
Received 18 October 2006; received in revised form 20 September 2007; accepted 16 October 2007
Available online 25 October 2007

Abstract
In this contribution, we consider mixed-integer nonlinear programming problems subject to differential-algebraic constraints. This class of
problems arises frequently in process design, and the particular case of integrated process and control system design is considered. Since these
problems are frequently non-convex, local optimization techniques usually fail to locate the global solution. Here, we propose a global optimization
algorithm, based on extensions of the metaheuristic Tabu Search, in order to solve this challenging class of problems in an efficient and robust
way. The ideas of the methodology are explained and, on the basis of two case studies, the performance of the approach is evaluated. The first
benchmark problem is a Wastewater Treatment Plant model [Alex, J., Bteau, J. F., Copp, J. B., Hellinga, C., Jeppsson, U., Marsili-Libelli, S., et al.
(1999). Benchmark for evaluating control strategies in wastewater treatment plants. In Proceedings of the ECC99 conference] for nitrogen removal
and the second case study is the well-known Tennessee Eastman Process [Downs, J. J., & Vogel, E. F. (1993). A plant-wide industrial process
control problem. Computers & Chemical Engineering, 17, 245-255]. Numerical experiments with our new method indicate that we can achieve an
improved performance in both cases. Additionally, our method outperforms several other recent competitive solvers for the two challenging case
studies considered.
2007 Elsevier Ltd. All rights reserved.

Keywords: Integrated process and control design; Mixed-integer nonlinear programming (MINLP); Metaheuristic; Tabu search; Tennessee Eastman plant

1. Introduction fore advisable, although one also should take into account the
degree of uncertainty associated with the process at the design
During the last decade, the importance of an integrated pro- level.
cess design approach, considering operability together with the More and more researchers are following the trend towards
economic issues, has been widely recognized. The aim is to considering the design and control aspects simultaneously. As
obtain profitable and operable process and control structures in a a result, a number of new methodologies has been developed
systematic way. Both the process design characteristics, control during the last years for addressing the solution of process
strategies, control structure and controllers tuning parameters design and process control problems. Sakizlis, Perkins, and
have to be selected optimally in order to minimize the total Pistikopoulos (2004) presented an overview of the state of the art.
cost of the system while satisfying a large number of feasi- Following this review the developed methods can be classified
bility constraints in the presence of time-varying disturbances. into two categories:
Consequently, the use of global optimization techniques is there-
The first category methods use a multi-objective approach.
In order to design an economically optimal process that can
This research was supported by the EU Marie-Curie Research Training operate in an efficient dynamic mode within an envelope
Network under project number MRTN-CT-2004-512233. around the nominal point, usually, the objective contains
Corresponding author. Tel.: +34 986 214473; fax: +34 986 292762.
of two parts. The first part represents the economic issues
E-mail address: julio@iim.csic.es (J.R. Banga).
URL: http://www.iim.csic.es/julio/ (J.R. Banga).
of the process, and the second one measures the dynamic
1 New address: University of Bayreuth, Department of Computer Science, performance, i.e. controllability, of the system. Since these
95440 Bayreuth, Germany. two parts are competing, weights have to be predefined to

0098-1354/$ see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2007.10.008
1878 O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891

represent the importance of the two objectives. In general,

Complete discretization on the dynamics and Generalized Benders decomposition is used to solve the

Control vector parameterization in combination with Outer approximation to treat the integers is used
Complete discretization on the dynamic system. The MIDO problem is transformed to a large mixed

Generalized Benders decomposition for treat integers. Uses a special integration gradient evaluation

CVP-GBD. Simplified master problem construction and no restriction to any integration or gradient
this is a hard task, since the quality of the of the solution
depends on these weights. A second drawback of these meth-

Complete discretization. The continuous variables are mapped in the discrete space. Then a
ods is that they also have problems to treat the dynamic of
the process in a systematic way. As a result, these meth-
ods often use a steady state model to determine the process
design.

integer nonlinear problem. This problem is solved using Outer Approximation

CVP-decomposition. New integer set obtained via screening model technique


The second category approaches use dynamic optimization
in order to find the most economic design. Instead of using
steady state models, the system operation is represented with

Complete discretization and a branch and bound method is used


a dynamic model. In this case a single objective which mea-

method that has benefits in the master sub-problem formulation


sures the economical performance can be used. The arising
optimization problem is a challenging mixed-integer dynamic

Generalized Benders decomposition to treat the integers


optimization (MIDO) problem.

Recently a number of algorithms were suggested for solving


the MIDO problem. The methods can be also categorized into
two groups. Table 1 shows an overview.

CVP-relax integers via stiff functions


Some methods reformulate the problem and as a result a
mixed integer nonlinear program (MINLP) has to be solved. The
MINLP is then solved by Outer Approximation, General Ben-

relaxation is performed
ders Decomposition or Branch and Bound frameworks. All these
methods are well-known for the ability to solve MINLPs. Draw-

evaluation method
backs of these methods are the facts that they either need to solve
relaxed problems, i.e. the integer constraints are dropped, or they

problem
solve a sequence of nonlinear problems (NLP) with fixed inte-
ger values. The latter situation implies a waste of computational
effort, if the integer values are far away from optimality. In this
sence our developed methodology is different, since we neither
solve relaxed problems nor do we solve NLPs with fixed inte-
ger values. The integer and the continuous variables are treated
simultaneously.
Avraam, Shah, and Pantelides (1998), Avraam, Shah, and Pantelides (1999), Balakrishna and

Mohideen, Perkins, and Pistikopoulos (1997) and Ross, Bansal, Perkins, and Pistikopoulos
The multimodal (non-convex) nature of MIDO problems has
Mohideen, Perkins, and Pistikopoulos (1996) and Dimitriadis and Pistikopoulos (1995)

been highlighted by, e.g. Schweiger and Floudas (1997) and


Bansal, Perkins, and Pistikopoulos (2000), among others. Con-
sequently, the use of global optimization (GO) techniques seem Bansal (2000) and Bansal, Sakizlis, Ross, Perkins, and Pistikopoulos (1996)

to be promising.
In the domain of deterministic GO methods, Esposito and
Floudas (2000) have presented approaches to solve dynamic
Sharif, Shah, and Pantelidis (1998), Schweiger and Floudas (1997)

optimization problems. This is indeed a very promising and pow-


erful approach, but restrictions may apply for the type of path
Biegler (1993) and Bahri, Bandoni, and Romagnoli (1997)

constraints which can be handled. Other groups (Singer, Bok, &


Barton, 2001; Papamichail & Adjiman, 2002) are also making
Methods for solving the MIDO (Sakizlis et al., 2004)

good progress in deterministic global optimization of dynamic


Fraga, Hagemann, Villagrana, and Bogle (2000)

systems, yet several issues regarding requirements (absence of


Samsatli, Papageorgiou, and Shah (1998)

noise, discontinuities, etc.) and computational performance are


still present.
Regarding stochastic GO methods, a number of researches
Schweiger and Floudas (1997)

have shown that they can locate the vicinity of global solu-
Control vector parameterization

tions for nonlinear dynamic problems with relative efficiency


Allgor and Barton (1999)

(Banga, Moles, & Alonso, 2003; Moles, Gutierrez, Alonso, &


Complete discretization

Androulakis (2000)

Banga, 2003; Sendin, Moles, Alonso, & Banga, 2004), but the
cost to pay is that global optimality can not be guaranteed.
However, in many practical situations these methods can be
satisfactory if they provide us with a good enough (often, the
(1998)
Table 1

best available) solution in modest computation times. Further-


more, stochastic methods do not require transformation of the
O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891 1879

original problem, which can be treated as a black box. Thus, Taking this into account the mathematical formulation is as
they can handle problems with complicated dynamics (e.g. follows:
discontinuities, non-smoothness, etc.). In summary, stochastic
methods do not provide guarantees that attraction basin of the min J(v, tf ),
v
global minimum has been located. On the other hand, deter-
ministic methods do provide guarantees, but they can not be s.t. f (x, x, p, v) = 0,
applied to realistic problems like the ones considered here due x(t0 ) = x0 ,
(1)
to the characteristics mentioned previously. However, stochas- h(x, p, v) = 0,
tic methods have shown, empirically, that they are very good
g(x, p, v) 0,
at locating the vicinity of the global solution in reasonable
times. vl v vu ,
In our case we present here a method based on extensions of
the metaheuristic Tabu Search (TS) which has been success- where v is the vector of decision variables, J is the objective
fully used in the area of operations research ((Glover & Laguna, function (often representing the costs) to minimize, x is the state
1997) for solving the integrated process and control system vector, p are some parameters, f is the set of differential and alge-
design problem. braic equality constraints describing the system dynamics and h
Tabu search has been recently reported to solve MINLP and g are possible equality and inequality path and/or point con-
problems in chemical engineering, such as process synthesis straints which express additional requirements for the process
(Linke & Kokossis, 2003; Lin & Miller, 2004), process design performance. The starting time for the simulation is denoted by
(Wang, Quan, & Xu, 1999), scheduling (Balasubramanian & t0 and the end, respectively, as tf . Consequently, x0 expresses
Grossmann, 2003) and computer-aided molecular design (Lin, the state variables at the beginning of the simulation. The lower
Chavali, Camarda, & Miller, 2005). The main differences of our and upper bounds for the decision variables are given with vl
approach with respect to these previous works are that (i) we and vu .
have improved the TS basic scheme regarding the generation A superstructure is developed that contains alternatives for
of new decision vectors (see detailed description below), and the process design. The common way to express alternatives
(ii) we have added a novel local solver (MISQP) to increase in the design is introducing binary variables. In this way the
the overall efficiency of the algorithm. This is particular impor- active state of a design alternative can be easily expressed. A
tant considering the computational cost of the dynamic systems value equal to one stands for an active alternative whereas in
embedded in the MINLPs that we are considering. the case of zero the alternative is inactive, i.e. not used. Thus,
This paper is structured as follows. The general statement of some of the elements of the decision variable vector v can be
the problem is presented in Section 2. Section 3 is dedicated restricted to integer values. As a result the problem formulated in
to the optimization methodology that we have developed. In (1) is a mixed-integer dynamic optimization (MIDO) problem.
Section 4 two challenging case studies are presented. The first It is quite known that nonlinear problems with a large number
benchmark problem is a Waste Water Treatment Plant model (see of integer variables are extremely difficult to solve to optimal-
Alex et al., 1999, and references therein) for nitrogen removal ity. Taking this into account the number of integer variables in
and the second case study is the well-known Tennessee Eastman the superstructure should be as small as possible. Engineering
Process (see Downs & Vogel, 1993). Finally, we present the knowledge and experience have to be used to limit the search
numerical results obtained by using our methodology and we space for the optimizer.
compare the performance of our approach with other competitive Although the resulting problem is very difficult to solve it is
modern solvers. worth investigating this kind of problems, since it is the only
way to obtain both, a process that has low cost and a process
that can be easily controlled. This approach is superior to all
2. The integrated process and control system design methods that only take one aspect into account. Reducing only
problem the cost might result in an instable process, while on the other
hand the best process from the controllability point of view can
In order to represent the interaction of process design and be very expensive.
control the formulation of the problem has to combine compo- As mentioned before, there are different approaches to solve
nents that express both design alternatives and the operability of this MIDO problem. Dynamic programming suffers from the
the system. In general a superstructure is developed that con- curse of dimensionality. Complete discretization, when applied
tains possible alternatives that one wants to consider. A key to problems which contain a significant number of dynamic
element that contributes significantly in the calculation of opti- states (like the examples considered here) results in a large
mal and sensible design solutions is the good definition of the number of decision variables, complicating the application of
design space and more specifically of the design superstruc- global optimization methods. As an alternative, we think that the
ture. Measuring the controllability of the process is achieved control parameterization approach, which results in a relatively
by introducing a system of differential and algebraic equa- small number of decision variables, facilitates the solution of the
tions that simulates the behaviour of the process under dynamic resulting MINLP using stochastic global optimization methods,
operation. like those based on Tabu search.
1880 O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891

The control parameterization technique parameterizes only First, let us summarize the basic idea of a Tabu search algo-
the control variables, so the decision vector will contain that rithm. The algorithm starts from an initial solution vk . For this
discretization information plus other selected time invariant current solution vk a set of neighbors is generated and the best
parameters. The optimization is carried out in the space of these among these neighbors is chosen to be the next iteration point
particular decision variables only. In our work we have followed vk+1 , even if the function value is worse than the one of the cur-
this technique, i.e. we discretize the control variables and obtain rent iterate vk . Allowing an increase in the objective function is
as a result a finite dimensional mixed-integer nonlinear pro- necessary to escape from a local minimum. To avoid cycling and
gramming (MINLP) problem with a dynamic system embedded, to guide the search into unexplored areas, some former visited
usually as a set of DAEs. points are set to be Tabu and so prohibited for some time. This
Since this resulting problem is frequently nonconvex, global procedure is repeated by starting from the new current point,
optimization (GO) methods are needed. As reviewed in the intro- until some stopping condition is fulfilled.
duction, deterministic GO methods for problems with DAEs The efficiency of the above mentioned procedure depends on
embedded can only handle (i) small problems and (ii) problems the choice of some key parameters, i.e., the length of the period a
which are differentiable and continuous. This is never met in point is set to be Tabu, and this is especially important in the case
realistic integrated design problems, so this is why these prob- of continuous variables. Since we do not know anything about
lems are so challenging, and why stochastic GO methods, like the topology of the objective functions, choosing right values
the one presented here, can be useful to overcome these chal- for these parameters is non trivial. Ideally, we would like to use
lenges. a procedure that is independent from such choices. This was
the main reason why we decided to use the basic concepts from
3. Description of the optimization method Battiti and Tecchiolli (1996). The authors proposed a TS algo-
rithm that is robust for any kind of functions and self-adjusting,
Many algorithms for global optimization use a local solver so that no parameters have to be set a priori. The components
to identify a local minimum by starting from an initial point, will be described now in more detail.
and in order to reach the global minimum, a special strategy
for deciding where to start the local solver is applied. For 3.1.1. Neighborhood
convex problems the local solver is able to find the global In the basic Tabu search for discrete optimization the neigh-
minimum. If the problem is not convex the global optimality borhood of an iteration point is built by all the direct neighbors
of the solution of the local solver cannot be guaranteed. In of this point. Since we have to handle large mixed integer prob-
this case one has to guide the local solver to the global opti- lems, it is impossible to consider all neighbors and it is even a
mum. This second component of the algorithm has to locate hard task to define exactly what a direct neighbor is. Thus, we
the attraction basin of the global minimum so that a run of have to restrict the neighborhood to some generated points. The
the local solver started in this basin will find the global min- idea is to subdivide the search space into regions that represents
imum. The approach developed here uses this strategy. As the basins of attraction and to perform the search over these regions.
global component, a procedure based on extensions of the Tabu As mentioned before a neighbor generating routine is used that
Search (TS) algorithm is applied. Regarding the local solver, is self-adjusting and independent from parameters that have to
we have used an special adaptation of a sequential quadratic be set a priori.
programming method for the mixed-integer case. The follow- The initial search region is specified by bounds on each inde-
ing subsection is dedicated to the description of the interaction pendent variable vi : vli vi vui , for i = 1, . . . , n. n is the
of the two different components and the basic ideas behind the number of independent variable and includes both continuous
proposed methodology, whereas the local solver by Exler and variables and integer variables. Initially, this search region is
Schittkowski (2007), called MISQP, is described in a separate subdivided into 2n equal-sized regions, obtained by dividing the
subsection. search space in half the range on each variable. This first guess
for the different regions is very crude but will be updated during
3.1. The hybrid strategyMITS the search. The regions are stored as a search tree and we perform
the search over the different nodes of the tree. In the beginning
TS is a metaheuristic originally developed by Glover (see the tree contains of 2n leaves representing the 2n regions. As
(Glover & Laguna, 1997). For the optimization of combinato- soon as two different local minima v and v are identified in
rial problems, TS has proved to be a very successful strategy. a region, the current region is subdivided into 2n equal-sized
During recent years, it has also been applied to the optimization smaller regions. If v and v belong to two different regions
of continuous problems. Here, an adaptation of TS to mixed- of the new partition, the splitting is terminated; otherwise the
integer nonlinear optimization problems, called Mixed-Integer splitting is applied to the region containig v and v until their
Tabu Search (MITS), is presented. This algorithm is an enhance- separation is obtained. Because of the splitting, a region can
ment of an approach proposed by Battiti and Tecchiolli (1996). contain at most one local minimum that was found so far. The
As Battiti and Tecchiolli, we also use a local solver to intensify regions will be stored internally as a tree and after some itera-
the search if necessary. The aim is to profit from the fast conver- tion of MITS the tree will be of varying depth in the different
gence of the local method. In this sense, the local solver MISQP regions, with regions of smaller sizes representing regions that
is integrated in the TS framework. require an intensification of the search. The leaves of the tree
O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891 1881

partition the initial search region, where the intersection of two


leaves is empty and the union of all leaves represents the initial
search space.
In each iteration MITS generates a set of points, called the
neighborhood N(vk ) of the current iterate vk . The iterate vk
lies in a unique region that can be identified by a well-defined
binary string B = [b11 , . . . , b1D , . . . , bn1 , . . . , bnD ], where D
represents the depth of the region in the tree. Initially, all region
are of depth one. After splitting a box, the new smaller regions
have a depth increased by one compared to their father regions,
i.e. the higher the value of D, the smaller the size of the region.
The length of the region edge along the ith coordinate is there-
fore equal to (vui vli )/2D . The position of the origin BiD of the
region containing the current iterate along the i-th coordinate is


D
bij
BiD = vli + (vui vli ) . (2)
2j
j=1 Fig. 1. Example for generated neighbors.

The neighborhood N(vk ) is a set of randomly generated points


in the neighbored regions. Depending on the depth of the cur- 3.1.2. Diversication vs. intensication
rent region up to 3 n neighbors are created randomly in the An important question arising in global optimization is when
corresponding regions. The number can vary according to the to intensify the search in a specified region and when to explore
size of the current region. The origins of the neighbored regions new regions. The risk is to waste a lot of time and evaluations
can be obtained as follows. In each coordinate up to two neigh- of the objective function in a region that does not contain the
bored regions are considered. Only the regions that do not violate global optimum, or to miss it if the search was not intensified.
the box constraints for the variables are evaluated. The origins Efficient metaheuristics provide a good compromise between
neighBD of these neighbored regions are obtained by changing diversification (exploration by global search) and intensification
only the ith coordinate according to the rule: (local search). In our approach the TS component has the purpose
of diversifying the search whereas the local solver intensifies the
vui vli search in a promising area found by the TS component and finds
neigh BiD = BiD . (3) the local minimum with high precision. This section specifies
2D
how this goal is accomplished.
The length of the neighbor region edge once again is (vui The diversification is done by generating the neighbors in
vli )/2D . In this way up to 2 n neighbors are created. Addi- the whole search space. In this manner every minimum can be
tionally, we start from the region of depth 1 that contains the reached. To enforce diversification we also prohibit recently vis-
current iterate that is represented by B1 = [b1 , . . . , bn ]. From ited regions. If the objective function or the constraint violation
this box we generate n regions where the origins are calculated for the infeasible case of the current iterate vk is less than all val-
as follows: ues of the generated neighbors, the local solver MISQP might
be started. To start MISQP:
vui vli
neigh Bil = Bil . (4)
2 J(vk ) J(v) for all v N(vk ) (6)
Fig. 1 shows an example for the generated trail points for a is a necessary condition, but it is not sufficient unless for the
two-dimensional problem. x is the current iterate and yi , i = first time a point in the region containing vk is locally optimal.
1, . . . , 6, are the generated trail points in the neighbored regions. In this case MISQP will always be executed. Otherwise, a new
The lower right region has been subdivided into 4 smaller regions run of the local solver has to be justified. Battiti and Tecchiolli
before. The lower left and the upper right region are of the initial (1996) apply some Bayesian stopping rules for multi-start global
size, but as one can see we generate two new trail points in each optimization to determine if a new run should be executed. We
of these regions. use the same strategy for MITS and so we summarize the idea
Among these neighbors the best one is chosen and the next in the following paragraph.
iteration starting from this point and the corresponding region. Let r > 1 be the number of times MISQP has been started in
For the unconstrained case that is the current region. An additional MISQP run should be started if
and only if there is a high probability of finding a new local min-
vk+1 N(vk ) : J(vk+1 ) = min J(v), (5) imum in the region. Starting MISQP in some parts of the region
v N(vk )
might lead the solver in a neighbored region. The region can be
where v is not Tabu, i.e. v does not lie in a region that is marked partitioned into W components, the attraction basins of the local
as Tabu. minima contained in the region and the possible basins that lead
1882 O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891

MISQP outside. The value of W is updated during the optimiza- When the described Tabu tenor changes are not sufficient and
tion depending on the information the local solver returns, i.e. cycles still occur an escape mechanism is needed. This mech-
termination inside a region or outside. It should be highlighted anism will force the search in new regions in a very drastic
that every region has its own value for W. MISQP is always exe- way. This is done by choosing randomly one of the generated
cuted if r W + 1. This means that so far every MISQP run neighbors that does not lie in a Tabu region. This is repeated
has detected a new attraction basin in the corresponding region. n times, where n is the number of variables.
In the other case, if r > W + 1 restarts have been performed
and W different components have been identified, we have to Since the number of generated neighbors depends on the size
justify a new run of the solver. Since at least one previous run of of the current region, respectively, the depth of the region in
MISQP has led to one minimum that had been located before. the tree, our prohibition rule has to take this into account. For
In this case MISQP is started only if the following equation is that reason we introduce the Tabu tenor fraction Tf [0, 1]. In
satisfied: other words a specified percentage of the neighbored regions is
(r W 1)(r + W) prohibited. In more detail, all neighbored regions that have been
rand > , (7) visited in the last:
r(r 1)

where rand is randomly generated number in the range [0,1] max(1, min(Tf n, n 2)) if D = 1,
T = (8)
(see, Battiti and Tecchiolli (1996)). In that way, the number of max(1, min(Tf 2n, 2n 2)) otherwise
MISQP runs can be kept down if the above estimate predicts a
small probability of finding a new local minimum, but a new run iterations are Tabu and cannot be visited. If D > 1 up to 3
is never completely prohibited for reasons of robustness. n neighbors can be generated, but since sometimes only 2
The initial starting point for MISQP is the current iterate vk n neighbors can be generated additional regions would lie
and the initial search region is the current region enlarged by outside the box constraints we use the formula Tf 2n, where
twice the range of the region. If v is a local minimum that was  denotes the largest integer value lower than Tf 2n. It might
found by MISQP, it is saved in a memory structure associated to happen that this value is too small but this is feasible because the
the region. If MISQP converges to a point outside the original escape mechanism will be activated, avoiding these cycles. In
region the solution will be the next starting point for the next iter- the normal situation, the list is proportional to Tf n (or Tf 2n),
ation; otherwise, MITS continues from the best neighbor found the max operator ensures that the list is at least one, the min
in this iteration. operator ensures that it is at most n 2 (or 2n 2). In this
As mentioned before, as soon as two different local minima v way, the last visited region is always prohibited, and at least two
and v are identified in a region, the current region is subdivided regions are available from a given point, so that the chosen move
into 2n equal-sized smaller regions. If v and v belong to two is influenced by the values of the objective function J for the
different regions of the new partition, the splitting is terminated; unconstrained case in the neighborhood.
otherwise the splitting is applied to the region containing v
and v until their separation is obtained. The local minima are 3.1.4. Aspiration criterion
associated with the new corresponding smaller regions. While the search is proceeding, several regions of the search
space are classified as Tabu. In some cases, the best neighbor
3.1.3. Tabu tenor update solution may lie in a Tabu area where its objective function
To guide the search into unexplored regions some of the pre- value is better than the current best value. In this situation the
vious visited regions are set to be Tabu, i.e. as soon as a region Tabu property can be invalidated and the point will be chosen.
contains the current iterate vk it is prohibited for the next iter- This feature is necessary to enforce faster convergence to a good
ation. Changing the Tabu tenor T the number of iterations a local minimum but it also might bias the search. If this happens
region is Tabu and cannot be revisited the algorithm can man- the escape mechanism will be started after some time and the
age the interaction between intensification and diversification. search is forced to leave this region and to explore a new one.
The following basic rules are used:
3.1.5. Stopping conditions
As soon as a region is visited again, the Tabu tenor is increased Defining a stopping criterion for GO algorithms is very dif-
and the algorithm is forced to start some diversification. The ficult. If the algorithm stops too early, the global optimum can
algorithm forces the search to choose a different neighbor for be missed. Otherwise, if it stops too late, computational effort
being the next iterate and new regions can be investigated. In will be wasted. Our procedure will stop if one of the following
this way cycling should be avoided. criteria is fulfilled:
When repetitions are absent for a sufficiently long period,
the Tabu tenor is reduced and the diversification will disap- A predefined maximum number of iterations is exceeded.
pear. In order to decide how long this period has to be the A given number of iterations without any further improvement
algorithm stores the average cycle length, i.e. the number of on the value of the objective function.
iteration it takes to revisit a region. If more iterations have A predefined maximum time has elapsed.
been taken than this average value, the Tabu tenor will be A predefined maximum number of function evaluations has
reduced. been reached.
O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891 1883

Since we have to handle computationally expensive models, the upper bounds on the variables. The number of equality con-
CPU time will be the bottleneck. Hence, in most cases we will straints h is expressed by me , the one for inequality constraints
stop our algorithm using the third criterion above. In our case g by mi . We define n = nc + ni the number of all variables
studies we deactived the second criterion since we wanted to and c = (h1 , . . . , hme , g1 , . . . , gmi )T the vector of constraints.
compare the performance of the different algorithm subject to Furthermore, we use the to declare a vector c with the
the maximum elapsed time respectively the maximum number following properties:
of function evaluations. Fig. 2 shows the flow chart of the MITS 
hj (x, y), j = 1, . . . , me ,
algorithm. The integration of MISQP in the framework MITS is cj (x, y) := . (10)
obvious in this figure. min(0, gj (x, y)), j = 1, . . . , mi .

The Lagrangian function plays a key role in nonlinear optimiza-


3.2. The local solver MISQP
tion and is defined as follows:
The purpose of this section is to give a short overview of 
me

the mathematical theory of the local solver that is activated by L(x, y, , ) : = J(x, y) j hj (x, y)
MITS. The solver called Mixed-Integer Sequential Quadratic j=1
Programming (MISQP) is a SQP Trust-Region method recently 
mi
developed by Exler and Schittkowski (2007). We will only sum- j gj (x, y), (11)
marize the basic ideas here, for further information please read j=1
the mentioned paper. We consider the general optimization prob-
where j , j = 1, . . . , me , are the Lagrangian multipliers for the
lem to minimize an objective function J subject to nonlinear
equality constraints and j , j = 1, . . . , mi , are the Lagrangian
equality and inequality constraints:
multipliers for the inequality constraints.
min J(x, y) MISQP is a Sequential Quadratic Programming method and
hj (x, y) = 0, j = 1, . . . , me , to enforce convergence a trust region method is used. We proceed
from the continuous trust region method of Yuan (1995) with
x Rnc , y Zni : gj (x, y) 0, j = 1, . . . , mi , (9) second order corrections. For the rest of this section subscripts
xl x xu , k denote the iteration. As a norm we use the -norm.
yl y yu , In each iteration k we have to solve the mixed integer
quadratic problem (MIQP):
where x denotes the vector of the continuous and y the vector of
the integer variables, nc is the number of continuous variables, 1 T
min k (d) =
d Hk d + J(xk , yk )T d
ni the number of integer variables, respectively. Once again, xl 2
d Rnc Zni : (12)
and yl denote the lower bounds, whereas xu and yu stand for + k ||(c(xk , yk )T d + c(xk , yk )) ||
||d|| k ,
where Hk denotes the approximation of the Hessian of the Lagra-
gian and dk is the solution to that MIQP in the k th iteration. The
step size is restricted to the trust region radius k . The constraint
violation is measured in the last term of k and is penalized with
parameter k . Here, denotes the first derivatives. Thus, J
is the gradient of the objective function, respectively, c is the
Jacobian of the constraints. The Hessian of the Lagrangian func-
tion is approximated by a quasi-Newton update formula subject
to the continuous and integer variables. In our implementation
the well-known BFGS-Formula is used.
It is not assumed that the MINLP is relaxable, i.e., that J(x, y),
hj (x, y), j = 1,. . ., me and gj (x, y), j = 1, . . . , mi , can be eval-
uated at any fractional parts of the integer variables. Thus, the
first derivatives at J(x, y) are approximated by the difference
formula:

J(x, y1 , . . . , yj + 1, . . . , yni ) J(x, y1 , . . . , yj 1, . . . , yni )


dy J(x, y) =
2
(13)

for j = 1, . . . , ni , at neighbored grid points. If either yj + 1 or


yj 1 violates a bound, we apply a non-symmetric difference
Fig. 2. MITSflow chart. formula. Similarly, dy hj (x, y) and dy gj (x, y) denote a difference
1884 O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891

formula for first derivatives for hj (x, y) and gj (x, y) computed In Algorithm 1 the formulation of the local solver MISQP is
at neighbored grid points. For the continuous variables the gra- presented. As a termination tolerance the parameter  is used.
dients are numerically approximated by a forward difference
formula. Algorithm 1 (MISQP).
After computing a step dk , one has to decide if the step is
accepted or not. Otherwise, convergence cannot be guaranteed.
We have to measure the improvement compared to a merit func- 0: Let (x1 , y1 ) Rnc Zni , 1 > 0, H1 Rnn positive defi-
tion. In our implementation we use the exact penalty function: nite, 1 > 0,  > 0, and let k := 1.
1: Solve subproblem (12) to get dk and the multiplier k and k .
P (x, y) := J(x, y) + ||c(x, y) || (14)
If k (0) k (dk ) <  and c(xk , yk ) < , then stop. Update
with a penalty parameter > 0. penalty parameter k .
The solution dk of the quadratic subproblem (12) is used to 2: Compute rk by (16). If rk > 0.75, goto Step 5. Solve sub-
compute the next iteration point: problem (18) to get dk and the corresponding multiplier k
and k , and compute rk by (20). Let k := k and := k . If
(xk+1 , yk+1 ) rk < 0.25, goto Step 3. If 0.9 < rk < 1.1, let k+1 := 2k ,

(xk , yk ) + dk if Pk ((xk , yk ) + dk ) Pk (xk , yk ) else k+1 := k and goto Step 6.
= (15) 3: If rk < 0.75, goto Step 4. Otherwise, compute J((xk , yk ) +
(xk , yk ) otherwise.
dk + dk ) and c((xk , yk ) + dk + dk ). If Pk ((xk , yk ) + dk +
A key role in the trust region algorithm is played by the prediction dk ) Pk ((xk , yk ) + dk ), goto Step 4. Calculate rk by (21)
of a new trust region radius for the next iteration. The trust region and let dk := dk + dk , rk := rk . If rk 0.75, goto Step 5. If
radius is adjusted according to the ratio of the actual and the rk 0.25, goto Step 6.
predicted reduction of the merit function: 4: Let k+1 := ||dk || /2 and goto Step 6.
5: If dk < k , then k+1 := k and goto Step 6. If rk >
Pk (xk , yk ) Pk ((xk , yk ) + dk ) 0.9, then k+1 := 4k , else k+1 := 2k .
rk := . (16)
k (0) k (dk ) 6: If rk > 0, goto Step 7. Otherwise, let (xk+1 , yk+1 ) :=
The trust region radius k has to be updated in order to enforce (xk , yk ), Hk+1 := Hk , increment k, and goto Step 1.
convergence. If rk is close to one or even greater than one, then 7: Define a new iterate (xk+1 , yk+1 ) := (xk , yk ) + dk , com-
k is enlarged and if rk is very small, k is decreased. If rk pute J(xk+1 , yk+1 ) and c(xk+1 , yk+1 ), update Hk+1 by the
remains in the intermediate range, k is not changed at all. More BFGS formula applied to dk and L(xk+1 , yk+1 , k , k )
formally, we use the same constants proposed by Yuan (1995), L(xk , yk , k , k ). Increment k and goto Step 1.
and set:

max[2k , 4||dk || ]
if rk > 0.9,
4. Case studies
k+1 = k if 0.1 rk 0.9, (17)


min[k /4, ||dk || /2] if 0 < rk < 0.1. Our algorithm, MITS, was implemented in Matlab. An inter-
face (software gateway) was also developed in order to call the
If, on the other hand, rk < 0, then k is decreased and we solve
local solver MISQP, which was implemented in FORTRAN (and
subproblem (12) again.
compiled as a dynamic link library). We have tested MITS on two
To increase the convergence rate a second quadratic problem
well-known and challenging benchmark problems: a wastewater
is used:
treatment plant (WWTP), and the Tennessee Eastman process
min k (d) (TEP).
d Rnc Zni : (18)
||d + dk || k , In optimal design problems, it is usually useful to perform
sensitivity analysis of the obtained optimal solutions. However,
where given the complexity of the models considered here, plus the fact
1 that they have noisy dynamics, standard codes for computing
k (d) : = (d + dk )T Hk (d + dk ) + J(xk , yk )T (d + dk )
2 sensitivities of dynamic systems (like, e.g. ODESSA) could not
be applied. Finite differences estimates could be used, although
+ k ||(c(xk , yk )T d + c((xk , yk ) + dk )) || . (19)
for noisy systems their reliability would be questionable unless
Let the solution be dk . In order to state the algorithm we define some sort of filtering was used, but this was out of the scope of
the following two equations: the present paper.
It should be noted that these benchmark design problems
k (0) k (dk ) were considered just as test cases for the investigation of
rk := rk + , (20)
k (0) k (dk ) the capabilities of the proposed global optimization technique.
and Therefore, questions regarding alternative formulations for these
problems (e.g. like considering other dynamic scenarios) were
Pk (xk , yk ) Pk ((xk , yk ) + dk + dk ) out of the scope of this paper and have not been investigated
rk := . (21)
k (0) k (dk ) here.
O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891 1885

Fig. 3. The wastewater treatment plant.

4.1. A wastewater treatment plant The problem is formulated as follows:


min J(v, tf ) = 1 EQ + 2 PE + 3 AE + 4 PSludge
The first benchmark problem that we consider is a wastewa- v
ter treatment plant (WWTP) for nitrogen removal. The WWTP +5 control
was developed by the COST 624 work group (19982004) s.t. x = f (x, x, p, v) = 0, (22)
and additional information can be found in Copp (2002) and
x(t0 ) = x0 ,
Alex et al. (1999). The layout of this benchmark plant as
shown in Fig. 3 combines nitrification with predenitrifica- vl v vu ,
tion by a five-compartment reactor with an anoxic zone. A
where f R150 denotes the system dynamics and v R12 Z
secondary settler separates the microbial culture from the liq-
is the vector of decision variables. The system dynamics are
uid being treated. A basic control strategy consisting of 2 PI
described by algebraic mass balance equations, ordinary differ-
controllers is proposed to test the benchmark. Its aim is to con-
ential equations for the biological processes in the bioreactors as
trol the dissolved oxygen level in the final compartment of the
defined by the ASM1-model (see Henze, Grady, Gujer, Marais,
reactor (AS Unit 5) by manipulation of the oxygen transfer,
& Matsuo, 1986), and the double-exponential settling velocity
and to control the nitrate level in the last anoxic compart-
function presented in Takacs, Patry, and Nolasco (1991) as a fair
ment (AS Unit 2) by manipulating the internal recycle flow
presentation of the settling process, with x R13 the state vector,
rate.
p Rp the system parameters and d the influent disturbance.
A Simulink implementation of the benchmark model by
In order to compare the controllability of different configu-
Jeppsson and Pons (2004) was used for the simulations. Each
rations, we define control as follows:
function evaluation consists of an initialization period of 100
days to achieve steady state, followed by a period of 14 1 ISE(N) (vk ) + 2 ISE(0) (vk )
control = , (23)
days of dry weather and a third period of 14 days of rainy 1 ISE(N) (v0 ) + 2 ISE(0) (v0 )
weather. Calculations of the controller performance criterion
where
are based on data from the last seven rain days. Each sim-  tf
ulation of this benchmark model takes a significant time on ISE() = ()2(.) d. (24)
a standard PC (about 60 s in a PC with Intel Pentium IV t0
3,2 GHz). The weights 1 and 2 are chosen such that the ISE(O) equals
The control performance was evaluated using the Inte- to the ISE(N) part when using the benchmark default set-
gral Square Error (ISE) as a criterion. Both the nitrate level tings v0 (i.e. the tuned PI-parameters) provided by the COST
and oxygen level controllers (further referred as N- and O- 624 (19982004). The values are 1 = 1/1001 and 2 =
controller, respectively) are optimized with respect to their 1000/1001. For the initial configuration v0 we obtain control =
controller parameters, that is, the gain K, integral time con- 1. If the controllability of a configuration is better than the one
stant i and anti-windup time constant t . Besides the controller at v0 the value of control will be smaller than one; otherwise, it
performance, the objective function also considers the follow- will be greater than one.
ing: The objective function is based on this formulated by
Vanrolleghem and Gillot (2002) which takes into account just
EQ. Effluent quality in kg pollution units/day includes number the economics of the process. We have added a controllability
of violations and percentage of time in violation. Limits for term to the cost function since previous tests revealed that just
each component are available. optimizing the criterion chosen by Vanrolleghem results in a
PE. Average pumping energy, in kWh/day poor controllability and high values for the pollutants concen-
AE. Average aeration energy, in kWh/day. tration. The weight for the control term is chosen so that its
PSludge . Sludge production, in m3 /day. value has the same order of magnitude as the economic term.
1886 O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891

Table 2 Table 5
Weights in the objective function Changes in the plant performance
Weight 1 2 3 4 5 Criterion Initial MITS Units
Value 2 1 1 3 1000
Objective function 35226 33537
value (J)
Effluent quality 9032 8356 kg poll units/day
Table 3
index (EQ)
Decision variables for the COST benchmark
Average aeration 7173 6787.6 kWh/day
Variable Description [vl vu ] energy (AE)
Average pumping 1919 2045.7 kWh/day
v (1) N-controller gain [100 1000] energy (PE)
v (2) N-controller integral time constant [0.0007 0.7] Sludge production 2357 2393.8 m3 /day
v (3) N-controller anti-windup time constant [0.0001 0.7] (PSludge )
v (4) O-controller gain [100 50,000] ISE(O) 1.2232 105 2.1779 104 (mg COD/l)2 day
v (5) O-controller integral time constant [0.01 1] ISE(N) 0.83345 0.46733 (mg N/l)2 day
v (6) O-controller anti-windup time constant [0.0001 0.07] Control 1 0.81
v (7) Aeration factor in reactor 1 [0 360] day1
v (8) Aeration factor in reactor 2 [0 360] day1
v (9) Aeration factor in reactor 3 [0 360] day1
v (10) Aeration factor in reactor 4 [0 360] day1 The new configuration obtained by our solver leads to sev-
v (11) Sludge purge [0 1844.6] m3 /day
eral changes, which are shown in Table 5. The performance has
v (12) Sludge recycle from settler [0 36,892] m3 /day
v (13) Feed layer in settler v(13) Z [1 10] improved for all the criteria except the average pumping energy
and the sludge production. This behavior can be explained taking
into account the multi-objective nature of the objective function
The exact values are listed in Table 2. It should be highlighted and the typical conflicting nature of the different criteria. As
that the values can be varied in order to change the weighting of already stated before, this behavior can be influenced by setting
different components, e.g., increasing the weight for the pump different weights. It is worth mentioning the change in the con-
energy the optimization will lead to a smaller value in the pump trollability: a significant improvement with respect to the default
energy. values was obtained, as shown in Table 5.
Table 3 lists all decision variables with a short description The performance of MITS was compared with the per-
and the upper and lower bounds. Boundaries on the decision formance of two other competitive modern solvers, namely
variables (vl and vu ) are chosen such that the process dynamics OQNLP and MINLPbb. OQNLP, based on the Scatter Search
would not show (exceptional) unstable behavior. metaheuristic, has been recently reported as one of the best
As mentioned before, the weights in the objective function global optimization solvers for black box problems (Neumaier,
are chosen with respect to the performance obtained with default Shcherbina, Huyer, & Vinko, 2005). The second solver,
values for the decision variables (i.e. the tuned PI-parameters) MINLPbb, has been developed by Leyffer (2001) and uses a
as provided by the COST 624 (19982004). The optimization branch and bound strategy to solve the MINLP. Fig. 4 shows
was started from these default values for the variables. Since the the convergence curves of the three different solvers. MITS and
evaluation of the objective function is costly (in the sense of time OQNLP were stopped after the time limit of two days exceeded.
consuming), we stopped MITS after exceeding a CPU time limit MINLPbb is a deterministic approach and stopped after the inter-
of 48 h using a standard PC. Table 4 shows the default values nal criterion was satisfied. In both categories objective function
(initial point for the optimizer) and the optimal values obtained value and convergence speed MITS clearly outperformed the
by MITS. other solvers.

Table 4
Result for the WWTP
Variable Initial point v0 Result MITS vMITS

v (1) 500 486.3


v (2) 0.001 0.000708
v (3) 0.0002 0.0001
v (4) 15000 16369.3
v (5) 0.05 0.0221
v (6) 0.03 0.0275
v (7) 0 0
v (8) 0 0
v (9) 10 h1 (240 day1 ) 224.8 day1
v (10) 10 h1 (240 day1 ) 224.8 day1
v (11) 385 m3 /day 333.1 m3 /day
v (12) 18446 m3 /day 16921 m3 /day
v (13) 5 7
Fig. 4. WWTPconvergence curves.
O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891 1887

4.2. The tennessee eastman process developed in a previous work (Antelo, Otero-Muras, Banga, &
Alonso, 2005) to derive robust decentralized controllers for the
Since the publication of the Tennessee Eastman process Tennessee Eastman Process. In this framework, the TEP is repre-
(TEP) example by Downs and Vogel (1993), it has been widely sented as a process network. Then, conceptual mass and energy
used in the literature as a benchmark due to its challenging inventory control loops for each node are designed first to guar-
properties from a control engineering point of view: it is highly antee that the states of the plant will remain on a convex invariant
nonlinear, open-loop unstable and it presents a large number of region, where the system will be passive and therefore input-
measured and manipulated variables which offer a wide set of output stability can be stated (Antelo et al., 2005). The next step
candidates for possible control strategies. The flowsheet for the is to realize the proposed conceptual inventory control loops
TEP is depicted in Fig. 5. Two products (G and H) are produced using the physical inputs-outputs of the process. Some extra con-
from four reactants (A, C, D and E). A further inert trace com- trol loops are needed to achieve the convergence of the intensive
ponent (B) and one byproduct (F) are present. The process units variables since the inventory control by itself does not ensure
consist of a continuous stirred tank reactor, a condenser, a flash the convergence of these variables to a desired operation point.
drum and a stripper. The gaseous reactants are fed to the reactor In some cases, the available degrees of freedom are not enough
where they are transformed into liquid products. The following to implement the complete control structure that ensures both
reactions take place in gas phase: extensive and intensive variables convergence to the reference
values. As a consequence, the setpoints of the inventory con-
A(g) + C(g) + D(g) G(l) , trollers can be used as new manipulated variables to complete
A(g) + C(g) + E(g) H(l) , the decentralized control design.
(25)
A(g) + E(g) H(l) , We explain the alternatives we introduced to extend the origi-
3D(g) 2F(l) . nal hierarchical control design proposed by Antelo et al. (2007).
Concerning the reactor level control loop in the original design,
These reactions are irreversible and exothermic with rates that its set point modifies the reference for the flow controller acting
depend on temperature through Arrhenius expressions and on over E feed. As an alternative for closing this loop, D feed is
the reactor gas phase concentration of the reactants. The reac- proposed as manipulated variable.
tion heat is removed from the reactor by a cooling bundle. The For the reactor pressure case, the original proposal by Antelo
products and the unreacted feeds pass through a cooler and, et al. (2007) considers the condenser cooling water flow as the
once condensed, they enter a vapour-liquid separator. The non- manipulated variable. By using this, the reactor pressure can be
condensed components recycle back to the reactor feed and the varied since the separator pressure and, as a consequence, the
condensed ones go to a product stripper in order to remove the recycle rate can be modified. It is at this point where the control
remaining reactants by stripping with feed stream. Products G over the vapor mass inventory in the separator by using the purge
and H are obtained in bottoms. The inert (B) and the byproduct rate is established in order to ensure that all inventories in the
(F) are mainly purged from the system as a vapour from the TEP will remain bounded and, therefore, inputoutput stability
vapour-liquid separator. is guaranteed (Antelo et al., 2007).
Recently, Antelo, Otero-Muras, Banga, and Alonso (2007) As an alternative, we are considering here a manipulated vari-
applied their systematic approach to a plant-wide control design able widely used in the literature for the reactor pressure control

Fig. 5. The tennessee eastman process flowsheet.


1888 O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891

loop: the purge flow. By modifying this, it is possible to regulate separator and the stripper volumes:
the separator pressure as well as the recycle flow, and there-
fore the reactor pressure. When this alternative is considered, an Preactor 3000 kPa,
extra loop controlling the separator temperature (energy inven- 2 m3 Vreactor 24 m3 ,
tory) by acting over the condenser coolant flow is defined. Note Treactor 175 C, (28)
that we are not considering other alternatives to control the reac-
1 m3 Vseparator 12 m3 ,
tor pressure as the A feed, that is a disturbance in the model, or
C feed. 1 m3 Vstripper 6 m3 .
In order to determine the best control alternative among The objective function proposed by Downs and Vogel (1993) in
the proposed ones, a new binary vector b is added to our the TEP definitions is based on the operating costs and can be
system dynamics. These 01 variables express which of the defined as follows:
four control strategies is being used, and they are defined as
follows: TC = PC PR + PrC PrR + CC CW + SC SR,
TC = 7.5973 $/kmol PC + 0.1434 $/kmol PrR (29)
b1 {0, 1} (E feed), + 0.0536 $/kWh CW + 0.0318 $/kg SC,
b2 {0, 1} (D feed),
(26) where TC are the total operating costs at the base case, PC and
b3 {0, 1} (condenser coolant), PR are the purge costs and purge flowrate, respectively. Analo-
b4 {0, 1} (purge flow). gously, PrC, CC and SC are the costs associated to the product
stream, compressor and steam, and PrR, CW and SR are the
product rate, the compressor work and the steam rate, respec-
Therefore, the original control design proposal by Antelo et al.
tively. Operating costs for this process are primarily determined
(2007) will be characterized by the vector b = (1, 0, 1, 0)T since
by the loss of raw materials (in the purge, in the product stream
it uses E Feed to control the reactor level and the condenser
and by means of the two side reactions). Economic costs for the
coolant flow to control the reactor pressure.
process are determined by summing the costs of the raw materi-
From all the exposed, the optimization problem consists now
als and the products leaving in the purge stream and the product
on solving the following MINLP of the form:
stream, and using an assigned cost to the amount of F formed.
The costs concerning the compressor work and the steam to the
minJ(x, v, b), stripper are also included. Note that the objective function used
v,b
in the MINLP formulation will be the mean of these operating
s.t. costs along the whole simulation time horizon. For this work,
f (x, x, p, v, b, t) = 0, this simulation time horizon was set to t = 10 h. This is enough
h(x, p, v, b) = 0, time for stabilization of the TEP.
g(x, p, v, b) 0, (27) After all these considerations concerning the objective func-
tion, the problem can be represented as an MINLP of the form
b1 + b2 = 1, (27):
b3 + b4 1 0,
v R36 , b {0, 1}4 :
vl v vu ,
bl b bu , min J(x, v, b) = total operating costs at base case
(30)
v0 0.5v0 v v0 + 0.5v0 .
where b {0, 1}4 is the vector of binary variables (01 vari- The lower and upper bound for the decision variables have been
ables) and v R36 are the continuous variables (the controller set to be the 50% of the initial value for the decision vector.
parameters). The lower and upper bounds for the binary vari- The reason for this selection is to avoid as much as possible the
ables will be of the form bl = (0, . . . , 0)T and bu = (1, . . . , 1)T . saturation problems that can be exhibited by the valves. These
It must be pointed out that we are considering that only one situations have been detected in preliminary dynamic simula-
of the two alternatives for each loop can be active at one time, tions when considering a value of 100% of v0 as bounds for
being necessary to introduce the additional linear constraints the decision vector.
b1 + b2 = 1. The linear constraint b3 + b4 1 0 ensure that Note that changes in the decision variables (v) will be trans-
at least one of the alternatives b3 or b4 is active. The rest of lated into variations in the states x that can even drive the system
the decision variables are connected to the tuning of the PI to shutdown due to the fact that one or more of the constraints
controllers. defined in (28) have been violated. More precisely, the reac-
Note that the MINLP is also subject to the dynamics (DAEs) tor volume, pressure and temperature are related with the pairs
of the system which are expressed by f in Eq. (27). The TEP has of decisions variables (gain-time constant) v(11) v(30), v(12)
171 DAEs (30 ODEs and 141 algebraic equations). The MINLP v(31) and v(16) v(35), respectively. Finally, the separator and
is also made up of the following constraints which are related the stripper volumes are linked with the pairs v(9) v(28) and
with the reactor pressure, temperature and volume, and with the v(10) v(29), respectively.
O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891 1889

Table 6
Result for the TEP
Decision variable Control loop Default values Result MITS

v(1) A Feed flow 0.001 1.4029e 3


v(2) D Feed flow 0.003 0.004043
v(3) E Feed flow 1.8e 10 9.6882e 11
v(4) C Feed flow 20 30.0
v(5) Condenser coolant 7e 7 9.8858e 7
v(6) Separator flow 4e 4 4.5727e 4
v(7) Stripper flow 0.004 0.004242
v(8) Production rate 3.2 4.8
v(9) Stripper level 0.02 0.021864
v(10) Separator level 0.05 0.048376
v(11) Reactor level 10 8.5024
v(12) Reactor pressure 0.0001 5e 5
v(13) %G in product 0.032 0.02811
v(14) %A in purge 0.0009 5.664e 4
v(15) Recycle rate 0.00125 0.001271
v(16) Reactor temperature 8 8.1636
v(17) Separator temperature 100 90.219
v(18) G /H Product ratio 32 21.762
v(19) G /H Product ratio 46 30.891
v(20) A Feed flow 1.6667e 5 1.584e 5
v(21) D Feed flow 1.6667e 5 1.4535e 5
v(22) E Feed flow 4.1667 3.1266
v(23) C Feed flow 0.1667 0.83334
v(24) Condenser coolant 4.1667 2.0834
v(25) Separator flow 1.6667e 5 1.42116e 5
v(26) Stripper flow 1.6667e 5 1.302e 5
v(27) Production rate 2 1.2995
v(28) Stripper level 0.3333 0.19213
v(29) Separator level 3.3333 1.6704
v(30) Reactor level 0.01667 0.011874
v(31) Reactor pressure 0.3333 0.18142
v(32) %G in product 1.6667 1.0927
v(33) %A in purge 9.3667 10.284
v(34) Recycle rate 25 30.338
v(35) Reactor temperature 0.125 0.06475
v(36) Separator temperature 8.3333 6.4701
b(1) Reactor level (E feed) 1 0
b(2) Reactor level (D feed) 0 1
b(3) Reactor pressure (condenser coolant) 1 0
b(4) Reactor pressure (purge rate) 0 1
Cost value ($/h) 156.843 84.289

In order to solve the MINLP (27) problem we used the


solvers MITS and OQNLP. The dynamic model for the proposed
thermodynamic-based control design has been also imple-
mented as a SIMULINK code by the Process Engineering Group
at IIM-CSIC. Both solvers were started from the same initial
point. As stopping criterion, we set the maximum of function
evaluations equal to 10000. Table 6 lists the default values (used
as initial point) and the best point located by MITS. The pairs
of
Note that the decision vector pairs v (1) v (20), v (2) v (21),
corresponds to the gain and time constants, respectively, of the
PI controllers employed to close the different loops presented in
Table 5. In other words, variables v (1) v (19) are the gains of
the PI controllers used to close the corresponding control loop,
while variables v (20) v (16) are the related time constants
of these controllers. Note that for the G/H Product Ratio there
are no time constants, since the associated controllers are only
proportional. Fig. 6. TEPconvergence curves.
1890 O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891

Note that for the case of the binary variables in Table 6, the to Computers & Chemical Engineering. In Presented in the 15th European
manipulated variable that has been chosen to close the loop is Symposium on Computer Aided Chemical Engineering (Proceedings of the
presented in brackets. The solution obtained by MITS for the ESCAPE-15), Vol. 20 (B) (pp. 11051110).
Antelo, L. T., Otero-Muras, I., Banga, J. R., & Alonso, A. A.
binary vector is b = (0, 1, 0, 1)T . This vector defines the new (2007). Hierarchical design of decentralized control structures for the
realization of the control loops for the pressure and the level in Tennessee Eastman process. Computers and Chemical Engineering,
the reactor by acting over the purge and the D feed, respectively. doi:10.1016/j.compchemeng.2007.10.021, in press.
Fig. 6 shows the convergence curves for both solvers. MITS Avraam, M., Shah, N., & Pantelides, C. (1998). Modelling and optimsation
clearly outperformed OQNLP, both regarding the final objective of general hybrid systems in the continuous time domain. Computers and
Chemical Engineering, 22(Suppl.), 221228.
function value and the convergence speed. Avraam, M., Shah, N., & Pantelides, C. (1999). A decomposition algorithm
for the optimisation of hybrid dynamic processes. Computers and Chemical
5. Conclusions Engineering, 23, 451454.
Bahri, P., Bandoni, J. A., & Romagnoli, J. A. (1997). Integrated flexibility and
controllability analysis in design of chemical processes. AIChE Journal, 43,
We have developed a hybrid algorithm for the optimization of 9971015.
the integrated process and control system design problem. We Balakrishna, S., & Biegler, L. (1993). A unified approach for the simulta-
have considered a mixed integer nonlinear program (MINLP) neous synthesis of reaction energy and separation system. Industrial and
formulation for this class of problems. Our novel hybrid strat- Engineering Chemistry Research, 32, 13721382.
egy, MITS, uses a combinatorial component, based on Tabu Balasubramanian, J., & Grossmann, I. E. (2003). Scheduling optimization under
uncertainty: an alternative approach. Computers and Chemical Engineering,
Search, to guide the search into promising areas, and a local 27, 469490.
solver, MISQP, which is activated to precisely approximate local Banga, J. R., Moles, C. G., & Alonso, A. A. (2003). Global optimization of
minima. A Matlab implementation of this technique was devel- bioprocesses using stochastic and hybrid methods. In C. A. Floudas & P.
oped, and its performance and robustness was tested on two M. Pardalos (Eds.), Frontiers in global optimization, nonconvex optimiza-
challenging benchmark problems: a wastewater treatment plant tion and its applications (pp. 4570). Kluwer Academic Publishers. ISBN
14020-76991
for nitrogen removal and the Tennessee Eastman Process. MITS Bansal, V. (2000). Analysis, design and control optimization of process systems
presented very good performance, and it was able to obtain under uncertainty. PhD dissertation, Imperial College of Science, Technol-
remarkable results on both case studies, clearly outperforming ogy and Medicine, London.
two selected modern MINLP solvers. Bansal, V., Perkins, J. D., & Pistikopoulos, E. N. (2000). Simultaneous design
The purpose of this paper was not to illustrate the perfor- and control optimisation under uncertainty. Computers and Chemical Engi-
neering, 24, 261.
mance of MITS for large-scale MINLP, but to show how it can Bansal, V., Sakizlis, V., Ross, R., Perkins, J. D., & Pistikopoulos, E. N. (1996).
solve complex integrated design problems where other state of New algorithms for mixed integer dynamic optimization. Computers and
the art solvers failed. The application of MITS to larger scale Chemical Engineering, 27, 647668.
problems is currently under research in our group. Initial tests Battiti, R., & Tecchiolli, G. (1996). The continuous Tabu search: Blending com-
performed with problems from standard collections have been binatorial optimization and stochastic search for global optimisation. Annals
of Operations Research, 63, 153188.
very promising so far, and they will be reported in future con- Copp, J. (2002). The COST simulation benchmark: Description and simula-
tributions. tor manual. Luxembourg: Office for Official Publications of the European
Community., p. 154.
Acknowledgements COST 624. (19982004). Optimal management of wastewater systems. Avail-
able: http://www.benchmarkwwtp.org/.
Dimitriadis, V., & Pistikopoulos, E. N. (1995). Flexibility analysis of
We acknowledge the support of the EU Marie-Curie Actions dynamic systems. Industrial & Engineering Chemistry Research, 34, 4451
program. In particular, we thank the EU PRISM project project 4462.
number MRTN-CT-2004-512233 Towards Knowledge-based Downs, J. J., & Vogel, E. F. (1993). A plant-wide industrial pro-
Processing Systems for the support. Author Jose A. Egea grate- cess control problem. Computers & Chemical Engineering, 17, 245
255.
fully acknowledges financial support (FPU fellowship) from the Esposito, W. R., & Floudas, C. A. (2000). Deterministic global optimization
Spanish Ministry of Education and Science. The authors wish in nonlinear optimal control problems. Journal Global Optimization, 17,
to thank Dr. Ulf Jeppsson (IEA, Lund University of Technology, 97126.
Sweden) for providing his MATLAB/SIMULINK implementa- Exler, O., & Schittkowksi, K. (2007). A trust region SQP algorithm for mixed-
tion of the COST Benchmark. integer nonlinear programming. Optimization Letters, 1(3), 269280.
Fraga, E., Hagemann, J., Villagrana, E., & Bogle, I. (2000). Incorporation of
dynamic behaviour in an autimated process synthesis system. Computers
References and Chemical Engineering, 24, 189194.
Glover, F., & Laguna, M. (1997). Tabu search. Kluwer Academic Publishers.
Alex, J., Bteau, J. F., Copp, J. B., Hellinga, C., Jeppsson, U., Marsili-Libelli, Henze, M., Grady Jr., C. P. L., Gujer, W., Marais, G. V. R., & Matsuo, T. (1986).
S., et al. (1999). Benchmark for evaluating control strategies in wastewater Activated sludge model no. 1 (Technical report no. 1). London, Great Brittain:
treatment plants. In Proceedings of the ECC99 conference. IAWQ.
Allgor, R., & Barton, P. (1999). Mixed-integer dynamic optimization. I. Problem Jeppsson, U., & Pons, M. N. (2004). The COST Benchmark simulation
formulation. Computers and Chemical Engineering, 23, 567584. modelcurrent state and future perspective. Control Engineering Practice,
Androulakis, I. (2000). Kinetic mechanism reduction based on an integer pro- 12, 299304.
gramming approach. AIChE Journal, 46, 361371. Leyffer, S. (2001). Integrating SQP and branch-and-bound for mixed integer
Antelo, L. T., Otero-Muras, I., Banga, J. R., & Alonso, A. A. (2005). A system- nonlinear programming. Computational Optimization and Applications, 18,
atic approach to plant-wide control based on thermodynamics. Submitted 259309.
O. Exler et al. / Computers and Chemical Engineering 32 (2008) 18771891 1891

Lin, B., Chavali, S., Camarda, K., & Miller, D. C. (2005). Computer-aided Samsatli, N., Papageorgiou, L., & Shah, N. (1998). Robustness metrics for
molecular design using Tabu search. Computers and Chemical Engineering, dynamic optimization models under parameter uncertainty. AIChE Journal,
29, 337347. 44, 19932006
Lin, B., & Miller, D. C. (2004). Solving heat exchanger network synthesis Schweiger, C. A., & Floudas, C. A. (1997). Interaction of design and control:
problems with Tabu search. Computers and Chemical Engineering, 28, optimization with dynamic models. In W. W. Hager & P. M. Pardalos (Eds.),
14511464. Optimal control: theory, algorithms, and applications. Kluwer Academic
Linke, P., & Kokossis, A. (2003). On the robust application of stochastic Publishers.
optimisation technology for the synthesis of reaction/separation systems. Sendin, O. H., Moles, C. G., Alonso, A. A., & Banga, J. R. (2004). Multi-
Computers and Chemical Engineering, 27, 733758. objective integrated design and control using stochastic global optimization
Mohideen, M., Perkins, J. D., & Pistikopoulos, E. N. (1996). Optimal design of methods. In M. Georgiados & P. Seferlis (Eds.), CACE book on Inte-
dynamic systems under uncertainty. AIChE J., 42, 22512272. gration of design and control (pp. 555581). Elsevier Science [ISBN
Mohideen, M., Perkins, J. D., & Pistikopoulos, E. N. (1997). Towards an effi- 0444-515577].
cient numerical procedure for mixed integer optimal control. Computers and Sharif, M., Shah, N., & Pantelides, C. (1998). On the design of multicompo-
Chemical Engineering, 21, 457462. nent batch distillation columns. Computers and Chemical Engineering, 22,
Moles, C. G., Gutierrez, G., Alonso, A. A., & Banga, J. R. (2003). Integrated 6976.
process design and control via global optimization: A wastewater treatment Singer, A. B., Bok, J. K., & Barton, P. I. (2001). Convex underestimators for
plant case study. Chemical Engineering Research and Design, 81, 507517. variational and optimal control problems. Computer Aided Chemical Engi-
Neumaier, A., Shcherbina, O., Huyer, W., & Vinko, T. (2005). A comparison of neering, 9, 767772.
complete global optimization solvers. Mathematical Programmes, Series B, Takacs, I., Patry, G. G., & Nolasco, D. (1991). A dynamic model of the
103, 335356. clarification-thickening process. Water Research, 25, 12631271.
Papamichail, I., & Adjiman, C. S. (2002). A Rigorous global optimization algo- Vanrolleghem, P. A., & Gillot, S. (2002). Robustness and economic measures
rithm for problems with ordinary differential equations. Journal of Global as control benchmark performance criteria. Water Science and Technology,
Optimization, 24, 133. 45, 117126.
Ross, R., Bansal, V., Perkins, J. D., & Pistikopoulos, E. N. (1998). A mixed- Yuan, Y.-X. (1995). On the convergence of a new trust region algorithm.
integer dynamic optimization approach to simultaneous design and control. Numerische Mathematik, 70, 515539.
In AIChE annual meeting. Miami Beach, Florida, US: American Institute of Wang, C., Quan, H., & Xu, X. (1999). Optimal design of multiproduct batch
Chemical Engineering. chemical processes using Tabu search. Computers and Chemical Engineer-
Sakizlis, V., Perkins, J. D., & Pistikopoulos, E. N. (2004). Recent advances in ing, 23, 427437.
optimization-based simultaneous process and control design. Computers &
Chemical Engineering, 28, 20692086.