Chemical Engineering Research and Design 189 (2023) 98–116

Available online at www.sciencedirect.com

Chemical Engineering Research and Design

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

Process optimization using a dynamic self-

adaptive constraint handling technique coupled to
]]
]]]]]]
]]

a Differential Evolution algorithm

J. Cortez-González a, A. Hernández-Aguirre b, R. Murrieta-Dueñas a,
⁎
R. Gutiérrez-Guerra c, , S. Hernández d, J.G. Segovia-Hernández d
a
Tecnológico Nacional de México, Instituto Tecnológico Superior de Irapuato, Departamento de Ingeniería Química
y Bioquímica, Carretera Irapuato-Silao km 12.5, C.P. 36821 Irapuato, Gto, Mexico
b
Centro de Investigación en Matemáticas, A.C., Departamento de Ciencias Computacionales, Callejón de Jalisco s/
n, C.P. 36240, Mineral de Valenciana, Guanajuato, Gto, Mexico
c
Universidad Tecnológica de León, Blvd. Universidad Tecnológica 225, Col San Carlos, C.P. 37670 León, Gto, Mexico
d
Universidad de Guanajuato, Campus Guanajuato, Departamento de Ingeniería Química, Noria Alta s/n, C.P.
36050 Guanajuato, Gto, Mexico

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

Article history: Nowadays, chemical processes are projected to obtain their best performance in energy and
Received 1 June 2022 water consumption, pollutant emissions and total annual costs, while still meeting quality of
Received in revised form products and good operational performance. These goals are accomplished through adequate
4 November 2022 optimization of the fitness function by manipulating the operational variables (decision
Accepted 7 November 2022 variables) of the process. However, a successful optimization process depends completely on
Available online 10 November 2022 the constraint handling established in the modeling of the process. The weighted summation
of constraint violations (weighting function technique, WF) is one of the most common ap-
Keywords: proaches for handling constraints in optimization problems. Nevertheless, in spite of this
Optimization technique yielding good results, in this work we show a novel self-adaptive constraint
Differential evolution handling technique (SA) based on a self-adaptation dynamic threshold and self-adaptation
Self-adaptive dynamic constraint (weight) factors. This technique deals with real and discrete variables and converts equality
handling constraints into inequality constraints through a dynamic threshold. Both penalization
Weighting constraint handling techniques (WF and SA) were, respectively, coupled to a Differential Evolution (DE) algorithm
Chemical processes to optimize some benchmark functions and chemical engineering optimization problems. In
addition, the rigorous model of a distillation train was optimized in Aspen One for the first
time with a self-adaptive constraint handling technique in chemical engineering. Although
both penalization techniques were coupled to the same DE algorithm and both cases were
run under the same conditions, the results show that the dynamic self-adaptive constraint
handling technique coupled to DE (DE-SA) achieves considerably better best-solutions than
the best-solutions obtained by the weighting function technique coupled to DE (DE-WF). In
addition, DE-SA led to substantial reductions of numerical effort in relation to DE-WF. These
conclusions are supported by statistical analysis of the results of 30 runs of the optimization
process for each constraint handling technique, for a distillation train.
© 2022 Institution of Chemical Engineers. Published by Elsevier Ltd. All rights reserved.

⁎
Corresponding author.
E-mail address:
[email protected] (R. Gutiérrez-Guerra).
https://doi.org/10.1016/j.cherd.2022.11.006
0263-8762/© 2022 Institution of Chemical Engineers. Published by Elsevier Ltd. All rights reserved.
 Chemical Engineering Research and Design 189 (2023) 98–116 99

1. Introduction adjust the weight factors. Several difficulties related to the

search for adequate penalty factors have been experienced
Sustainable development demands chemical processes able as they require parameters for the cooling model which is
to transform raw materials into final products using optimal very sensitive to the performance of the algorithm. In addi-
amounts of energy and water, minimizing waste and redu- tion, some methods eliminate equality constraints by redu-
cing emissions of pollutants into the environment. cing the dimension of the search space (Dasgupta and
These goals can be tracked by implementing robust de- Michalewicz, 1997) or introduce random key encoding to
sign and optimization strategies. However, due to the com- maintain the feasibility of the solution (Asafuddoula et al.,
plex nature of the chemical processes, their mathematical 2015). Another approach attempts to move infeasible solu-
models are nonlinear, multivariable and constrained. Thus, a tions into feasible regions (Liepins and Vose, 1990). Un-
key concern is the constraint handling technique used to fortunately, this approach is problem-dependent and
keep chemical processes under optimal design and opera- computationally expensive in cases of complex optimization
tional conditions. Constraint handling is not commonly problems.
emphasized when deterministic or evolutionary algorithms On the other hand, if an infeasible solution can be easily
(EA) are implemented to optimize chemical processes. fixed, repair algorithms are usually a better choice as they
For the evolutionary algorithms the constraint handling use non-dominance-based selection in which constraints are
techniques has been classified intro five categories: methods handled as additional objectives (Coello and Mezura-Montes,
based on penalty functions, methods based on special re- 2002; Mezura-Montes and Coello, 2011).
presentations and operators, methods based on repair algo- Other authors have incorporated a gradient-based repair
rithms, methods based on separation of objectives and algorithm with a dominance-based selection scheme
constraints, and other hybrid methods (Coello, 2002). The (Kheawhom, 2010). Repair algorithms are usually a good
penalty function is the most common approach in the evo- choice if an infeasible solution can be easily fixed
lutionary algorithm to handle constraints. In this method, (Boukouvala and Ierapetritou, 2014; Yu et al., 2019).
the constrained optimization problem is transforms into an Some stochastic global optimization algorithms have
unconstrained optimization problem that includes pena- been used to solve constrained optimization problems by
lizing the infeasible points and the degree of constraints means of the feasibility approach (Chanthasuwannasin et al.,
violation (Kheawhom, 2010). The methods based on special 2017), or by finding a feasible point using the gradient of
representations and operators, consist in reduces the di- constraints at an infeasible point (Takahama and Sakai,
mension and simplifying the shape of the search space 2006). Other authors have proposed re-formulating (Zhang
through of a change of representation in the problem and Rangaiah, 2012), hiding (Na et al., 2017) or eliminating
(Dasgupta and Michalewicz, 1997). In the methods based on (Yiqing et al., 2007) equality constraints to solve non-convex
repair algorithms, attempts to move infeasible solutions into nonlinear programming (NLP)/mixed integer nonlinear pro-
the feasible region (Riche and Haftka, 1997). Typically, gramming (MINLP) problems using evolutionary algorithms
heuristic rules are used to guide the repair process. In the such as Differential Evolution, Particle Swarm Optimization,
other approaches separately handle objective functions and Genetic Algorithm and self-adaptive algorithms.
constraints (Separation of constraints and objectives). These Several authors in chemical engineering have used static
approaches require several extra parameters and the con- penalty functions to solve optimization problems with in-
straints are can handled as additional objectives (Coello and equality constraints, due to their effectiveness and ease of
Mezura-Montes 2002). Finally, the hybrid methods normally implementation. NLP and MINLP problems have been solved
are coupled with a numerical optimization approach to using a Differential Evolution algorithm and Tabu list
handle constraints in evolutionary algorithm (Adeli and (Srinivas and Rangaiah, 2007) coupled to the penalty function
Cheng, 1994). method. Improvements to the determination of penalty
In fact, a weighting penalization function is the most parameters (weight factors) in the weight functions have also
applied approach to handle constraints during the optimi- been proposed (Sreepathi and Rangaiah, 2017). Finally, an-
zation process (which triggers some difficulties in selection other approach consists of a constraint handling method
of a convenient penalty factor). The weighting penalization utilizing process characteristics to formulate the weight
function can be traced back to the first numerical algorithms function (Austbø et al., 2013; Austbø and Gundersen, 2016). In
used to solve constrained optimization problems. They be- these implementations, the results show that the penalty
came the de factor approach due to their broad applicability. function method with an adequate value for the penalty
The penalization concept is simple to grasp: for every vio- parameter provides better optimal solutions.
lated constraint, a penalty factor is applied to the optimiza- Other penalty approaches use several constraint handling
tion function; therefore, functions that do not violate the techniques to enhance the convergence speed and popula-
constraints have the best function value. The trick with this tion diversity. For engineering design problems, an im-
approach is to find the weight (factor) of each penalty such plementation supported by the gradient repair method and
that larger penalizations should correspond to larger con- ranking by constraint fitness has been analyzed (Zahara and
straint violations. Kao, 2009). One technique involves the assignment of a
Static penalties retain their weight value throughout the random sequence of constraints to the solution, the number
optimization process. Dynamic penalties change their value of satisfied constraints and the violation measure
during the process. There are methods based on dynamic (Asafuddoula et al., 2015). Another technique uses a repair
penalties in which the penalty increases over generations or algorithm based on the gradient information derived from
that employ information from the search to adjust the pe- the equality constraints to deal with infeasible solutions
nalization (Joines and Houck, 1994; Kazarlis and Petridis, through dominance-based selection (Kheawhom, 2010). Re-
1998; Dadios and Ashraf, 2006). Other methods use a simu- cently, Yu et al. (2019) proposed a mechanism which trans-
lated annealing algorithm (Michalewicz and Attia, 1994) to forms the equality constraints into inequality constraints,
100 Chemical Engineering Research and Design 189 (2023) 98–116

combined with seven criteria to compare feasible and in- agent subject to the absorbent compositions. Furthermore,
feasible solutions. Their results showed that the constraint we present the implementation of the proposed technique
handling scheme is very important for identifying feasible for the rigorous optimization of a distillation train through
solutions early and reducing computational cost. evaluation of the objective function with Aspen One. This
Despite the sophistication of the constraint handling problem is non convex, non-linear, multivariable subject are
techniques reported in the literature, they have only been equality and inequality constraints.
applied for continuous search spaces and explicit objective The paper is organized as follows. In Section 2, we in-
(fitness) functions (in other words, the objective function is a troduce the new algorithm coupled to a Differential Evolu-
simplified mathematical function of the problem). In several tion algorithm and implementation of the optimization
research's, Aspen has been used as an evaluator of the fit- process. In Section 3, we show the results of the benchmark
ness function in processes with constraints, for example in- functions, the explicit chemical engineering problems and,
tensified distillation systems (Vazquez–Castillo et al., 2009), a finally, the optimization and statistical analysis of the dis-
crude distillation system (More et al., 2010), a natural gas li- tillation train. In Section 4, we discuss the results and in
quefaction processes (Austbø and Gundersen, 2016), an ex- Section 5 we present our conclusions.
tractive distillation (Franke, 2019) and recently a multiple-
interconnection process (Lyu et al., 2022). In these cases the 2. Optimization algorithm and constraint
optimization problem is subject a set of constraints. handling
In this paper we present a self-adaptive constraint
handling technique coupled to a Differential Evolution algo- Constrained optimization problems are characterized by the
rithm. This approach computes the weight factors based on fact that the solution must be found inside the feasible re-
the magnitude of the violation and the number of constraints gion which is delimited by the equality and/or inequality
not satisfied for each individual during the optimization constraints. This means that the search for solutions favors
process. In other words, when the amount of violation is those individuals that better meet the constraints and that
large, a strong penalization is applied, and when the viola- also maximize or minimize the objective function.
tion is smaller, a smaller penalization is used. Additionally, Generally, a constrained nonlinear optimization problem
for each individual, if the number of violated constraints is is expressed as follows:
high, the penalization factor is strong, and it is weak when Find vector
the number of violated constraints is small. In our proposal,
the dynamic threshold searches for feasible individuals and Z = (z1, …, zD ), Z RD
it is only updated when the whole population accomplishes
To minimize fitness function
every constraint. This dynamic threshold approach requires
the transformation of equality constraints into inequality f (Z )
constraints.
Subject to equality constraints
We present validation of the proposed technique with five
benchmark problems and three typical chemical engineering hj (Z ) = 0 j = 1, …, n
problems that have solutions in continuous spaces.
Subject to inequality constraints
The benchmark problems selected are representative of
the nature of chemical engineering problems. The first pro- gk (Z ) 0 k = 1, …, m
blem corresponds to an NLP problem, this problem has five
variables and three equality constraints and taken from Deb And subject to boundary constraints
et al. (2000). The second problem is a NLP problem with four zi(L) zi zi(U ) i = 1, …, D
variables, two linear inequality constraints and three non-
linear equality constraints. This problem has been previously Thus, the optimization goal is to find a feasible vector Z ,
studied in Yiqing et al. (2007). The third problem is taken composed of D dimensions, to minimize the fitness function
from Diwekar et al. (1992), this MINLP problem and it con- f (Z ) accomplishing the equality hk (Z ) and inequality con-
tains five variables, four inequality constraints and five straints gj (Z ). Evolutionary algorithms need to be coupled
equality constraints and represents two reactors to minimize to a constraint handling technique that allows them to find
the cost of producing a desired product. The fourth problem feasible solutions in a constrained search space. The chosen
is a non-convex MINLP problem also studied in Summanwar technique plays an important role in the quality of the so-
et al. (2002), it has three binary and two continuous, in this lutions delivered and its computation time.
problem the two equality constraints are the sources of Techniques that favor the feasible solutions are named
nonconvexities for this example. The fifth problem con- overpenalized and may lead to a fast convergence to a fea-
sidered a heat exchanger network synthesis problem re- sible solution but it may be suboptimal. Premature con-
quiring minimization of the total cost, has formulated it as a vergence means that the search space is not adequately
one variable problem with 12 constraints proposed by explored during the initial optimization process (Ben, 2016).
Summanwar et al. (2002). The typical chemical engineering On the other hand, techniques that favor the infeasible so-
problems selected are a heat exchanger network, a reactor lutions are named underpenalized and can lead to a slow
network and an absorption process. These case studies were convergence to the feasible zone or even no feasible solution
taken from Edgar et al. (2001). For heat exchanger network at all.
the total annual cost was minimize and is subject to a loga- A weighting function (WF ) technique is the more usual
rithmic mean temperature difference. In the reaction net- way to handle constraints in chemical engineering.
work the objective function was minimization of the total Nonetheless, finding the correct weights is an issue which we
volume, subject to an overall conversion. Finally, in the ab- propose to overcome via self-adaptation. In this technique,
sorption process, we optimize the flowrate of the absorber f (Z ) is the fitness function penalized by the weighted sum of
 Chemical Engineering Research and Design 189 (2023) 98–116 101

constraint violations, wj are the factors that we need to find constraints and is progressively reduced throughout the op-
and f (Z ) is the new optimization function. timization process, additionally transforms equality con-
m straints into inequality constraints, as shown below:
f ´(Z ) = f (Z ) + wj max(0, Gj Z)
j=1 htarget
j (z ) hreached
j (z ) = 0

In this case, for the WF technique to handle the con-

is treated as:
straints we used two sets of penalties (Leboreiro and
Acevedo, 2004). A major penalty was applied when the in- |htarget
j (z) hreached
j (z)| >
dividual was infeasible in evaluator (non-convergence eva-
luator), while a successful simulation was just slightly Where: htarget
j is the set point of the constraint (target value)
penalized. One is used to penalize infeasible individuals and hreached
j (z) is the value of the constraint reached, for each
(w1 = 1000 ) and the other penalizes each time the constraints individual. In our approach, the constraint is considered as
are not met (w2 = 100 ). Otherwise, in both cases the weight accomplished when |htarget j (z) hreached
j (z )| is within the in-
factor value is 0. According to the experiments realized, these terval ( ; ) and viola j = 0 . The threshold reduces when the
values perform well for most problems. whole population accomplishes all constraints, violatotal = 0 ,
Larger or smaller values of wj lead to the overpenalization f constraints = 0 and f pen = f total .
or underpenalization described. In chemical engineering, The progressive reduction of epsilon is defined by the
some authors have proposed constraint handling techniques following equation:
based on the WF approach that can be applied to problems in
update = c*
continuous spaces (Teh and Rangaiah, 2003; Babu and
Angira, 2006; Kheawhom, 2010; Zhang and Rangaiah, 2012). Where: is the current value of the dynamic threshold, c is
For this reason, it is necessary to design a constraint hand- the progressive reduction factor and update is the new value of
ling technique capable of coupling with evolutionary algo- the dynamic threshold. The factor c is adjusted between 0.5
rithms to optimize complex problems with mixed variables and 0.9. For each problem, this value must be empirically
and highly nonlinear models. adjusted. According to the experiments realized, a value of
0.8 performs well for most problems. In this implementation,
2.1. Self-adaptive constraint handling technique the initial value of is 0.5.
An important aspect considered is when violatotal > 0 ; in this
Generally, the optimization of chemical engineering pro- case, we calculated the self-adaptation of penalty factors that
blems uses design constraints in discrete and continuous depend on the degree of deviation of the constraint between
search spaces are multivariable and multimodal; in addition, the reached value ( hreached
j (z)) and the target value (htarget
j (z )).
the objective function is highly nonlinear and non-convex For each individual, the penalty factors for each constraint are
subject to nonlinear equality and inequality constraints. calculated according to the following expression:
Therefore, robust constraint handling techniques, which wconstraints
j = b*(htarget
j (z ) hreached
j (z ))2
consider the information of non-feasible and feasible in-
dividuals to favor the optimization process, are required. Notice that the penalty factor is computed using the re-
Next, we present a constraint handling technique that can sidual value (htarget
j (z ) hreached
j (z)) raised to the power 2 mul-
be coupled to evolutionary algorithms, in this case tiplied by coefficient b, for each equality constraint, wconstraints
j .
Differential Evolution (DE). This approach includes four dif- The power 2, guarantees positive values and multiplying by b
ferent penalization criteria and transforms equality con- enables the handling of similar orders of magnitude in order
straints into inequality constraints through a dynamic to establish significance in the penalty method. When the
threshold involving self-adaptive factors. Fig. 1 shows the residuals are large, a strong penalization is applied. For a
general self-adaptive (SA) penalization approach. small violation, the penalization is weak (see Fig. 1, lines
15–28). This step allows the best fitness function routes to be
A. Infeasible individuals powered and no promissory search zones to be segregated by
the algorithm.
In the first step, individuals are tested for feasibility and
the death penalty applied for non-convergence in the eva- C. Self-adaptive penalties
luator (violaconvergence = 1). This penalty in only applied to non-
feasible individuals, and the worst value ( f max ) is assigned to In addition to the magnitude of the violation, our tech-
the fitness function ( f fitness) and a poor value (s) is assigned at nique also involves counting the number of unmet con-
all equality constraints (hj ). Otherwise, the individual is straints, violatotal , and computes f num_pen . The penalization is
evaluated to obtain f total and the constraint values (hreached
j (z)). carried out as a function of the number of constraint viola-
This step is shown in lines 2–14 in Fig. 1. tions, violatotal ; the fitness function is penalized by a factor
proportional to the amount the constraint is violated,
B. Self-adaptive dynamic threshold wnum_pen . For each individual, if the number of violations is
small the penalization is weak, and it is strong when the
In the second step, the main objective is to search areas of number of violations is high (see Fig. 1, lines 29–36).
the feasible region through relaxing and stiffening of the Static penalties and self-adaptive penalties are applied in
constraints, to determine the directions of movement our proposal (Homaifar et al., 1994; Coello, 1999). In this kind
through the dynamic threshold . of penalization, the user defines several levels of violation
In other approaches, this uses as level comparison be- and penalty coefficients (penalty factors) for each constraint.
tween the aptitudes of the parent and the child (Takahama Thus, the penalty coefficient increases as higher levels of
and Sakai, 2010). In this proposal is used to relax-stiff constraint violations are experienced.
102 Chemical Engineering Research and Design 189 (2023) 98–116

Fig. 1 – General self-adaptive penalization scheme.

These penalties are represented by integer factors, in- D. Constraint on design variables
itialized to 0 and incremented by 1 for each constraint of
the problem that is violated, regardless of the amount of In several problems, there are design variables that are
violation. Notice that only the number of constraints vio- bounded by a maximum value mmax , usually defined heur-
lated is considered in the variable violatotal , but not the istically. This value is not a constraint but will be considered
magnitude by which each constraint is violated. Through as an important part of the nature of the problem. We apply
this combination of penalties, the evolutionary algorithm this type of penalty to guide the search towards solutions
has enough information about how several constraints are that find the appropriate values of this design variable, ziselect .
violated, but also about the amount of constraints un- This kind of penalization is applied depending on the
accomplished. nature of the optimization problem. For instance, if mini-
mization of the fitness function is pursued, the penalty
 Chemical Engineering Research and Design 189 (2023) 98–116 103

method decreases the final fitness function value only if all altered, the only thing that is favored is the intensification in
restrictions lie within the dynamic threshold, ε. The fitness feasible regions that minimize the thermal load and also the
function using the penalty is f num_pen . So, the reduction of the total number of stages.
fitness function depends on the relationship between ziselect In the area of evolutionary computation, when devel-
and mmax . This favors intensification of the promising area, oping constrained optimization algorithms it is necessary
since this relationship is only between 0 and 1. In the last to design a constraint handling technique that allows the
step, the evolutionary algorithm returns f fitness (see Fig. 1, algorithm, during the optimization process, to find the
lines 37–50). When any restriction is not met, this stage does feasible zone. In this sense, the techniques that have been
not apply and f fitness = f num_pen . recently developed consider sophisticated elements such
This technique allows to establish superiority over fea- as the implementation of as a repair algorithm based on
sible points and to partially order them in order to guide the the gradient information as rank-based selective pressure
search towards individuals that decrease the objective strategy (Fister et al., 2021) or basic concepts such as
function and that also implicitly minimize some of the di- weight function. However, an important point of discus-
mensions (design variables) of the individual. To explain the sion in the area of evolutionary computation is the de-
implementation of this technique we will use as an example termination of additional parameters corresponding to the
the case study of the distillation train. It is important to constraint handling technique. Traditionally, the tuning of
mention that in the optimization of distillation columns we these parameters depends on the nature of the optimiza-
know that a decrease in the thermal load, increases the tion problems to be solved. That is, the technique is
number of stages of the column, which is why some authors adapted to the characteristics of the problem, which limits
have proposed to perform a multi-objective optimization its generalization, as expressed by Fister et al. (2021), this is
with stochastic algorithms (Vázquez-Ojeda et al., 2013; because the algorithms are adapted to the type of functions
Gómez - Castro et al., 2015; Medina - Herrera et al., 2017; to be optimized.
Palma - Barrera et al., 2019; Contreras - Zarazúa et al., 2021). In particular, our technique uses 3 user-defined para-
Our implementation allows using a uniobjective algo- meters: b, ε and c. These parameters depend on the nature of
rithm, where the constraints can be additional objectives in the problem to be solved.
the optimization process and it is possible to find zones that
meet the decrease of both the heat duty and the number of • Parameter b is assigned a constant value during the whole
total stages of the separation system. optimization process, which guarantees that the order of
In the case study of the distillation train, we assign as magnitude of fconstraints is equal to that of f fitness. This
"constraint on desing variable" the total number of NT stages parameter allows the penalty to be significant according to
of each individual. NT is the sum of the number of stages of the degree of constraint violation. That is, if it violates a
all the columns involved in the separation scheme. For this lot, it penalizes strongly, otherwise it penalizes softly.
last stage, it is necessary to set a heuristic maximum value • The parameter ε defined as dynamic threshold is depen-
related to the feasibility of construction of the distillation dent on the nature of the problem. The function of this
column (mmax). The value of mmax is used in two stages, in parameter is to relaxed the constraints at the beginning of
the first one to compare it with NT and in the second one to the optimization process, since it will accept as feasible
calculate the penalty factor wz . In the comparison between individuals those that at that moment comply with the
mmax with NT, when the value of NT is less than that of mmax relaxed constraint. However, during the optimization
then the fitness function should be penalized. The penalty process, the value of ε is updated, turning the constraints
factor is determined from the quotient between NT and mmax , rigid, leading the search towards feasible zones.
and a value less than 1 is obtained. Therefore, when the fit- • The parameter c is a constant value throughout the opti-
ness function is penalized, the numerical value will be fur- mization process and corresponds to a progressive re-
ther decreased. In a minimization problem, the individuals duction factor of ε. Values of this parameter are between 0
that have the highest probability of being chosen are those and 1. Values very close to 1 cause epsilon to reduce
with a lower value of fitness function. So, implicitly, with this slowly. Values close to 0 increase the rate of epsilon re-
strategy the selection pressure increases for those in- duction. This parameter helps to control the selection
dividuals that minimize the objective function and also im- pressure, avoiding premature convergence.
plicitly decrease the number of total stages of the distillation
train. Thus, we define a superiority of feasible points, i.e., The weighting factors between ftotal and fconstraints , which do
when the child faces the parent, the one with the lowest not require initial values, are calculated from the parameters
fitness function value and therefore the lowest number of b and These parameters must be modified depending on the
total stages in the separation system will be chosen. This nature of the problem. According to our experience these
strategy intensifies the search for the feasible region where proposed values obtain optimized values in more complex
the objective function is minimized and also reduces the configurations, and it is not necessary to modify them.
number of design stages of the separation scheme. Our However, it is necessary to perform the tuning of the para-
proposal can include more than one design variable in the meters of the proposed technique along with the chosen
constraints. This allows to include some additional objective optimization algorithm.
as a constraint in a uniobjective algorithm. Recently, we have It should be noted that the technique has no dependence
implemented this technique to decrease the pressure and on the optimization algorithm, so it can be coupled to any
the number of total stages in oscillating pressure distillation algorithm of the user-define considering the tuning of the
systems, obtaining quite favorable results. parameters of both the optimization algorithm and the
It is important to mention that in this penalty technique constraint handling technique.
none of the design variables (dimensions) of the individual is
104 Chemical Engineering Research and Design 189 (2023) 98–116

Fig. 2 – Flowsheet of Differential Evolution algorithm.

2.2. Differential evolution algorithm for elitism, if after “n” number of generations the fitness
function of elite individual remains constant. Other case is if
The classic DE algorithm consists of four stages: initializa- the epsilon value decreases by more than 10–3 of all or part of
tion, mutation, crossover and selection (Storn and Price, the total population. Finally when some parameter of the
1997). As can be seen in Fig. 2, once the initial population is evolutionary algorithm.
generated, the optimization loop is performed. Notice that In DE, the initial population of individuals is randomly
the mutation, crossover and selection stages are carried out generated within the boundary constraint of each design
to generate the new individuals of each generation. Here- variable (search space). Each individual i contains char-
after, these stages are performed continuously until the acteristic decision variables D and is called a decision vector
specified stopping criterion is satisfied. In this case, the Zi, G . All individuals Npob in each generation G are selected as
number of function evaluations is established as the stop- parents and each of them is processed as described below: in
ping criterion. This criterion is the most commonly used in general, in the mutation process, three individuals are ran-
evolutionary algorithm literature (Zielinski and Laur, 2008). domly chosen as parents (Zr1, G, Zr2, G, Zr3, G ), except the parent
In this case, we selected the number of iterations as stopping in the processing (Zi, G ).
criteria, however, it is possible to implement a stop criterion
 Chemical Engineering Research and Design 189 (2023) 98–116 105

The base vector, Zr3, G , is the first individual and the re- function, f total . The evaluated individuals are analyzed
maining two are used to obtain the residual (Zr1, G Zr2, G ) that according to the constraints violated, for each constraint
is then scaled by the scale factor F and added to the base handling technique. This information is added to the pe-
vector; this means that the resulting vector is recombined nalized fitness function f fitness which is used by the algo-
with the parent. The recombination probability is controlled rithm to generate the new population G + 1. This process
by the crossover factor CR . The result of the crossover pro- is repeated until the stopping criterion is reached.
cess produces a trial vector, vi, G . Finally, in the survivor se- • The distillation train optimization process requires the
lection operator, the trial vector is accepted in the generation implementation of an interface between MATLAB, Excel
G + 1 if the trial vector is better than the parent (see Fig. 2). and Aspen One©. The DE algorithm is coded in MATLAB,
The mutation process is performed based on the dis- the SA constraint handling technique in Excel and the
tribution of the solutions in the current population. In this fitness function evaluator in Aspen One©. The process
way, search directions and possible step sizes depend on the optimization consists of the DE algorithm generating the
location of the individuals selected to calculate the mutation population, Npob, for each generation, G . Each individual,
values. The CR parameter controls the influence of the parent Zi, G , is sent to Aspen One to simulate the distillation train
in the generation of the offspring. Higher values mean less to obtain the objective function, f total , and the purities and
influence of the parent. The F parameter scales the influence recoveries for each component xipur and xirec . This in-
of the set of pairs of solutions selected to calculate the mu- formation is used to determine the fitness function pe-
tation value. The mutation and crossover steps together with nalization f fitness. The fitness function value is returned to
the selection step constitute one generation or iteration of MATLAB to generate the new population, G + 1. This pro-
the DE algorithm. The procedure is repeated until the spe- cess is repeated until the stop criterion is met. A detailed
cified stopping criterion is satisfied. description of the implementation of the SA constraint
handling technique for the distillation train is in Appendix
2.3. Constrained optimization procedure A. Notice that the constraints are not treated within the
Aspen simulator, but in Excel where the implementation
In the constrained optimization process, the DE algorithm of the constraint handling technique was performed. In
was coupled to our self-adaptive constraint handling tech- this case, our optimization proposal is based on the eva-
nique (DE-SA) and weighting penalty function (DE-WF), for luation of the objective function through a sequential
comparison of results. In this optimization process, the DE modular simulator Aspen Plus, so the optimization com-
algorithm requires the individuals to be physically feasible in puting time considers three stages: the evaluation of the
all generations. An individual being physically feasible im- objective function (Aspen Plus), the evaluation of the
plies that it satisfies the problem and that it is possible to constraint handling technique (Excel) and the optimiza-
calculate the objective function through an evaluator. tion process of the DE algorithm (MatLab). Of these three,
For the first generation, in this work the population was the most time-consuming stage is the one performed by
randomly generated as follows: Aspen Plus due to the time required to achieve con-
vergence of the simulation and obtain the value of the
zl(L) + randomj [0, 1] (zl(U ) zl(L) ) objective function, as stated by Kiss et al. (2012). In our
if z i, l, G + 1 < zl(L) or zi, l, G + 1 > zl(U ) case optimization computing time corresponds to the
zi, l, G + 1 = evaluation time of each individual and is 4 s
zi, l, G + 1 = zl(L) + |zi, l, G + 1 zl(L) |
zi, l, G + 1 = zl(U ) |zi, l, G + 1 zl(U ) |
3. Case studies and results
where:
i = 1, …, Npob l = 1, …, D This work presents an SA technique based on a dynamic
threshold and self-adaptation. This constraint handling
Through this approach, the set of variables that comprises
technique has been coupled to a DE algorithm (DE-SA). We
each individual, i, is kept within the boundaries established.
compared the performance of the SA technique with results
In this case, N represents the size of the population for each
obtained by weight function penalization (DE-WF) for the
generation and zl(L) and zl(U ) represent the lower and upper
same problems. The DE algorithm is coupled to the two
limits of the variables or dimensions of problem, respec-
techniques. Thus, the performance of DE-SA versus DE-WF is
tively. This allows to replace the individuals with variables
particularly influenced by the penalization technique instead
out boundaries for individuals that if accomplishing them.
of the DE algorithm. The values of the parameters for the DE
Explicit and implicit fitness function problems are treated
algorithm were: CR = 0.8 and F = 0.85. These are typical va-
here. Explicit fitness functions are benchmark functions and
lues used in the literature (Wong and Dong, 2005). The DE-SA
typical chemical engineering problems. The distillation train
and DE-WF algorithms were run 30 times for each problem.
case is an implicit fitness function. The constrained optimi-
In the distillation train case, we performed a statistical
zation procedure is different in each case, as described
comparison of the results.
below:
We optimized five benchmark functions with constrained
in continuous space, proposed in CEC 2006 and solved by
• The constrained optimization process for the case of the
Kheawhom (2010), to validate our SA technique. In addition,
benchmark functions selected for CEC 2006 and chemical
we analyzed three explicit chemical engineering problems
engineering functions was implemented in MATLAB. The
proposed by Edgar et al. (2001). These case studies represent
DE algorithm generates the population Npob in each gen-
complex optimization problems, from constrained NLP to
eration G and this is evaluated to obtain the objective
constrained non-convex MINLP.
106 Chemical Engineering Research and Design 189 (2023) 98–116

Table 1 – Benchmark functions for constraint optimization.

Finally, the SA technique was implemented in a more 3.1. Optimization of benchmark functions
complex search space. This technique was implemented to
optimize a multicomponent distillation scheme based on In this case, the five benchmark functions shown in Table 1
MESH equations. This mathematical model is highly non- were optimized. These benchmark functions have been
linear, non-convex and multimodal, with continuous and widely used as optimization tests (Kheawhom, 2010). The
discrete variables subject to equality and inequality con- first benchmark represents an NLP problem. The second
straints. In this problem, the fitness functions were eval- benchmark is the modeling of an NLP problem with four
uated in Aspen One using a computer with the following variables, two linear inequality constraints and three non-
characteristics Core i7 processor, 2.5MGz and 8 GB RAM. linear equalities. The third benchmark is an MINLP problem.
The fourth benchmark represents a non-convex MINLP
 Chemical Engineering Research and Design 189 (2023) 98–116 107

These case studies were taken from Edgar et al. (2001) and
Table 2 – Optimization results.
formulated as minimization problems. Minimization of the
Test Function Fitness function (f fitness ) total annual cost was established as the fitness function in
type
Kheawhom Weighting Self-adaptive
the heat exchanger network optimization problem, subject to
(2010) function (DE-SA) a logarithmic mean temperature difference (LMTD ). The fit-
(DE-WF) ness function for the reaction network optimization problem
was minimization of the total volume (VT ), subject to an
1 NLP 0.0539498 0.0646498 0.0539498
overall conversion of 0.9 for XA3. The flowrate minimization
2 NLP 5126.5 5127.42 5126.5
3 MINLP 99.245209 99.5209 99.245209 of the absorber agent (solvent) was defined as the fitness
4 Non- 7.66718 7.928 7.66718 function of the absorption process optimization problem
convex subject to the absorbent compositions y1 and y2 being within
MINLP the range of y0 to y3.
5 MINLP − 1.923098 − 1.9897 − 1.923098 These case studies are composed of continuous search
spaces and nonlinear fitness functions. In addition, these op-
timization problems are subject to equality and inequality
problem, while the fifth case is formulated as an MINLP constraints. Likewise, both case study 2 and case study 3 have
problem that contains three integer variables and five con- an important dependence between the optimization variables
tinuous variables and several equality and inequality con- due to reaction rates and equilibrium ratios, respectively. The
straints. In this case, 20,000 function evaluations and 100 formulation of the optimization problems is shown in Table 3.
individuals per generation were used in the optimization In this optimization, a population of 60 individuals per gen-
process. These problems are representative of the nature of eration and 3000 function evaluations were used.
some chemical engineering problems.
3.2.1. Results
3.1.1. Results Fig. 3 shows the optimization results for both cases, DE-SA
Table 2 shows the optimization results for the five bench- (blue line) and DE-WF (red line), and they are compared with
mark functions. The optimization based on dynamic self- the optimal value determined by Edgar et al. (2001) using a
adaptive constraint handling (DE-SA) obtained the same deterministic method (green line).
optimal values of f fitness reported in the literature According to the behavior depicted in Fig. 3, it was observed
(Kheawhom, 2010) for all cases evaluated. The values de- that DE-SA found the optimal solution in around 650 function
termined with the weight function penalization technique evaluations, while DE-WF obtained the optimal value in about
(DE-WF) are larger than the values reported, particularly in 1800 function evaluations. Thus, it is evident that this perfor-
the last two problems that correspond to MINLP problems. mance of DE-SA significantly reduces the computational effort
SA and WF use the same stop criterion as Kheawhom (2010) by more than 60%. In fact, DE-SA obtains the best solution in
to validate the results. between 1 and 3 min, while obtaining the optimal value with
Kheawhom (2010) performed optimization of the bench- DE-WF takes between 3 and 5 min. So, at least 50% reductions
mark functions using a DE algorithm coupled with a repair of computation times are obtained.
algorithm for constraint handling. Comparing the results Table 4 shows the optimization results determined using
obtained using DE-SA, DE-WF and those obtained by DE-SA and DE-WF and the corresponding results reported by
Kheawhom (2010), a large difference in the number of func- Edgar et al. (2001). According to the results for the three case
tion evaluations used to optimize the benchmark functions studies, the solutions obtained using the SA technique were
was observed. In fact, the optimization performed by DE-SA very similar to the optimal values reported previously by
required 20,000 function evaluations, which corresponds to Edgar et al. (2001), while those obtained using the WF tech-
using a population of 100 individuals and 200 generations. nique experienced a larger deviation in the optimal value
Comparatively, the optimization performed by Kheawhom and greater computing effort was required.
(2010) required 2,000,000 function evaluations, which results The statistical analysis showed that the SA algorithm is
from a population of 1000 individuals and 2000 generations. robust and efficient since the same optimal value was de-
On the other hand, the optimal value of the fitness func- termined in the 30 runs achieved. So, it is evident that the SA
tion was obtained in generation 35 in the case of DE-SA technique is a worthy tool since it is able to deal with opti-
whereas the optimal value obtained by Kheawhom (2010) mization problems subject to equality and inequality con-
was determined in generation 120. Therefore, a considerable straints and interdependence on the optimization variables,
reduction of numerical effort and computing time was re- as the separation schemes.
quired by DE-SA. This demonstrates the robustness and ef-
ficiency of the constraint handling technique to deal with 3.3. Distillation train
problems with different levels of complexity, from NLP pro-
blems to highly nonlinear and multivariable optimization In this case study, a distillation process (Fig. 4) is used to split
problems with strong interactions between the variables, a mixture made of four lineal aliphatic hydrocarbons: n-bu-
such as non-convex MINLP with inequality and equality tane, n-pentane, n-hexane and n-octane. A flow rate of 45.36
constraints. kmol/h is introduced in the first column as a saturated liquid.
The feed composition is as follows: 0.45 for intermediate
3.2. Optimization of typical chemical engineering components and 0.05 for the other components. The design
problems specifications for the purity and recovery for each compo-
nent were established as 0.987 for A, 0.98 for intermediate
The following case studies were optimized: a heat exchanger components and 0.986 for the last component. The design
network, a reaction network and an absorption process. pressure for the separation was chosen to ensure using
108 Chemical Engineering Research and Design 189 (2023) 98–116

Table 3 – Typical chemical engineering case studies.

pur
cooling water in the condensers. The phase equilibrium for h1, j (Z ) = x target j = 1, …, Mconstraints
the liquid phase was determined using the Chao–Seader
rec
model, since the system is a hydrocarbon mixture to be se- h2,j (Z ) = x target j = 1, …, Mconstraints
parated at low or medium pressure (Aspen Plus 13.0, 2007).
And subject to boundary constraints
The optimization problem is the following:
Find vector zi(L) zi zi(U ) i = 1, …, D

Z = (NT, B1, NF, B1, NT, B2, NF, B2, NT, B3, NF, B3, RRB1, RRB2, RRB3 ) 2 NF,Bl NT, Bl l = 1, …, Bcolumns
NS1, NS2 NF
To minimize fitness function

Q = f (Z ) xipur = x 0pur

Subject to constraint function xirec = x 0rec

 Chemical Engineering Research and Design 189 (2023) 98–116 109

Besides that, it was determined that most individuals

concentrate between 20 and 40 GW/y whereas the best fit-
ness function values represent a heat duty of 20 GW/y.
However, in spite of lots of individuals lying around this
optimum value, most of them do not meet all constraints.
The results generated in the optimization using DE-SA are
shown in Fig. 5b). In this case, it is evident that from the first
generations (300 function evaluations) that the DE-SA algo-
rithm obtains feasible solutions for a value of the threshold.
These initial solutions were feasible as the threshold is re-
latively large initially. Nevertheless, notice that from about
300–6500 function evaluations, the penalty function is larger
due to the threshold value (ε) reducing as optimization pro-
gresses. This means that the population must have larger
Fig. 3 – General performance of the optimization process in
purities and recoveries than the first individuals, which in-
a typical chemical engineering case. creases energy consumption. Also observe that as the
threshold reduces, the DE-SA algorithm has difficulty in, first,
finding a feasible zone and, second, achieving stabilization in
where: NT is the total number of stages, NF is the feed stage the feasible zone, increasing the fitness function.
and RR is the reflux ratio. NT and NF represent discrete vari- Nonetheless, notice that after 6500 function evaluations,
ables, while RR represents a continuous variable. the DE algorithm was capable of identifying and remaining in
Consequently, there are nine dimensions, which represent a feasible zone. Thus, the robustness of the algorithm led to
six discrete variables and three continuous variables. the generation of feasible solutions that meet purity and
Minimization of the total heat duty (Q ) in the reboiler is recovery targets with a set of design variables that allow
the fitness function of the problem. Notice that Q is a func- minimization of the total heat duty (fitness function). This
tion of the vector Z . The design specification for purity (purity behavior was determined in an experience of 30 experi-
pur
constraint) is defined as x target and that for recovery (recovery ments (runs).
rec
constraint) is given by x target . As before, in order to conduct a On the other hand, Table 5 shows the best five solutions,
comparative analysis, optimization of the problem was per- from 20,000 total function evaluations, obtained by DE-WF.
formed using DE-SA and DE-WF, respectively. In this case, a Notice that the best solution has a fitness function of
population of 100 individuals and 200 generations were used. 21.5846 GW/y and the worst individual has a fitness function
The CPU time of the optimization process corresponds to the of 32.6248 GW/y. In addition, the results in Table 5 evidence
total time required to evaluate 20,000 individuals and is ap- that all constraints are accomplished, but all purities and
proximately 22 h. A high percentage of the time corresponds recoveries are above the target value. Therefore, in terms of
to the long convergence time required by the Aspen Plus si- design, high values of purity and recovery mean high energy
mulator that ensures rigorous process simulation results consumption, which save concordance with the fitness
(Vázquez-Ojeda et al., 2013). function values.Table 6.

3.3.1. Results 3.3.2. Comparative analysis

As shown in Fig. 5a), the results indicate that DE-WF has Based on the previous analysis, it is clear that the dynamic
some difficulty in exploring effectively and determining op- self-adaptive constraint handling presented in this paper
timal values in the feasible search zone. This behavior is improves the performance of the DE algorithm through a
produced due to the weight function penalization of feasible more intensive exploration of the search space. It led to
individuals (designs) having a similar order of magnitude as better solutions, less numerical effort and a shorter compu-
infeasible individuals. In other words, the weight penaliza- tation time than the other constraint handling technique.
tions were not enough to improve the optimization process. These results are a consequence of the dynamic behavior of
The wide dispersion of the results reflects this. the threshold during the optimization process. It is

Table 4 – Objective function results for typical chemical engineering cases.

Test Objective function Edgar et al. (2001) Our technique

Weighting function Self-adaptive

(DE-WF) (DE-SA)

1 0.65 2428.022 2428.025 2428.019

CAT = ( 350
na )
w1 cp1 (T1S T1E )
U * TML,1
+
w2 cp2 (T2S T2E )
U * TML,2
min
0.005 * 8500
+ lvap
(w1 cp1 (T1F T1S ))

2 Go (yo y1 ) Go (y1 y2 ) 55.154 56.789 55.154

LT = x1
+ x2
min
Go (y2 y3 )
+ x3

3 minVT =
QCAF (XA1 XA2)
+
QCAF (XA2 XA3) 814.648 815.742 814.646
rA1 rA2
110 Chemical Engineering Research and Design 189 (2023) 98–116

Fig. 4 – Distillation train.

Fig. 5 – Optimization of distillation train, DE-WF vs DE-SA.

important to point out that the degree of relaxation and re- assumed to be normally distributed. The null hypothesis in
duction of the threshold was determined using information the bootstrap signed-rank test is that no significant differ-
of feasible and non-feasible individuals. ence exists between the performance of DE-SA and that of
On the other hand, taking into account that the major DE-WF. The alternative hypothesis is that there is a sig-
compromise in a distillation column is given by the total heat nificant difference in the performance of DE-SA compared to
duty and total number of stages, it is important to point out its competitor. In the following expressions, H0 represents
that the designs found by DE-SA are better for both NT and Q the null hypothesis and H1 is the alternative hypothesis.
than those determined by DE-WF. In fact, the best design
H0 : µDE SA = µDE WF
obtained with DE-SA requires 56% less energy (9.4257 GW/y
vs 21.5846 GW/y) and 26% fewer stages (82 stages vs 111
H1: µDE SA µDE WF
stages) than the best design determined by DE-WF. Hence,
the best solution obtained by DE-SA provides the best com- where: µDE SA and µDE WF represent the average of the fitness
promise between energy consumption and total number of function values for DE-SA and DE-WF, respectively. The
stages for this distillation train. bootstrap signed-rank test works by comparing two para-
In addition, in order to conduct a more rigorous statistical meters, called α and p-value. α is defined as significance level
analysis, DE-SA and DE-WF results were compared using a and it is considered as the reference parameter in this ap-
nonparametric statistical hypothesis test called the boot- proach. In the literature, a standard value of 5% (0.05) is as-
strap signed-rank test (Hesterberg, 2011). This test is used signed for α. Hence, the null hypothesis is rejected if the p-
when the average of the optimal fitness values can be value is less than the significance level. Outcomes obtained
 Chemical Engineering Research and Design 189 (2023) 98–116 111

Table 5 – Best five designs found by DE-WF.

Design 1 2 3 4 5

Total heat 21.5846 26.8670 30.4748 31.9111 32.62484

duty,
Q (GW /y)
Column 1, B1
NT 49 41 44 27 50
NF 33 21 18 19 31
RR 15.9506 14.2624 19.6461 20 20
Column 1, B2
NT 23 28 44 28 31
NF 11 16 25 18 16
RR 8.6047 20 15.6566 10.7606 8.4996
Column 1, B3
NT 39 42 25 25 50
NF 8 22 7 14 49
RR 7.5152 1.9634 8.1922 13.6602 16.2600
Purity
xApur 1.0000 0.9999 0.9999 0.9999 1.0000
xBpur 0.9920 0.9920 0.9920 0.9920 0.9920
xCpur 0.9921 0.9921 0.9843 0.9921 0.9891
xDpur 1 0.9999 0.9798 0.9999 0.9703
Recovery
xrec
A 1.0000 0.9999 0.9999 0.9999 1.0000
xBrec 0.9920 0.9920 0.9920 0.9920 0.9920
xCrec 0.9921 0.9921 0.9843 0.9921 0.9891
xDrec 1 0.9999 0.9798 0.9999 0.9703

Table 6 – Best five designs found by DE-SA.

Total heat 9.4257 10.4795 10.9875 11.5112 12.0309
duty,
Q (GW /y)

Column 1, B1
NT 18 41 32 44 42
NF 11 12 11 14 25
RR 20 14.4178 16.9542 16.8805 17.4485
Column 1, B2
NT 34 47 38 27 36
NF 21 27 5 17 12
RR 3.9457 4.6335 5.1058 3.2294 5.0529
Column 1, B3
NT 30 35 38 15 21
NF 19 28 12 7 5
RR 1.4821 2.1354 1.9390 3.9878 2.7984
Purity
xApur 0.9905 0.9990 0.9992 1.0000 1.0000
xBpur 0.9910 0.9920 0.9837 0.9909 0.9921
xCpur 0.9920 0.9916 0.9839 0.9904 0.9922
xDpur 0.9993 0.9946 1.0000 0.9943 0.9999
Recovery
xrec
A 0.9905 0.9990 0.9992 1.0000 1.0000
xBrec 0.9910 0.9920 0.9837 0.9909 0.9921
xCrec 0.9920 0.9916 0.9839 0.9904 0.9922
xDrec 0.9993 0.9946 1.0000 0.9943 0.9999

through the bootstrap signed-rank test at α = 0.05 confirm obtained through a resampling with replacement of the best
that the performance of DE-SA is better than that of DE-WF. fitness function values. These fitness function values were
This statement is also supported by the results obtained by determined by carrying out 30 experiments (runs) for each
applying other statistical tests such as the median and var- algorithm (30 runs for DE-SA and 30 runs for DE-WF). In this
iance. case, a resampling rate of 80% was applied. Thus, from 30
On the other hand, with the aim to support earlier results, total individuals, 24 individuals were randomly chosen and
a confidence interval was determined. This interval was the mean was computed. The resampling was performed
112 Chemical Engineering Research and Design 189 (2023) 98–116

Fig. 6 – Statistical analysis comparing DE-SA and DE-WF.

50,000 times, obtaining the histogram shown in Fig. 6a) persion in the experiments achieved using DE-SA in relation
and b). to the experiments carried out using DE-WF. This can be
The results obtained by DE-SA showed that in about 95% observed through the size of the box with respect to the y
of essays (runs) the fitness function values of the best in- axis. In addition, observe that the DE-SA experiments show
dividuals lie between 10.3422 and 15.6312 GW/y, with an that even the solutions obtained in the last quartile are a
average value of 12.8605 GW/y and a median of 12.8450 GW/y. substantial improvement on those obtained with DE-WF.
Meanwhile, the range of the fitness function values for the Regarding CPU time, DE-SA and DE-WF require approxi-
best individuals determined using DE-WF varies between mately 22 h per run to obtain the optimal designs. There is no
23.9886 and 41.6991 GW/y, the average being 31.1890 GW/y substantial difference between both techniques, since a high
and the median 31.1071 GW/y. These intervals are re- percentage corresponds to the evaluation of the fitness
presented through the red bars shown in Fig. 6a) and b), function in Aspen Plus. Authors such as Vazquez–Castillo
where a normal distribution is evidenced. Through the be- et al. (2009) optimize Intensified distillation systems in qua-
havior shown, it is clear than the optimization performance ternary mixtures through genetic algorithms, evaluating
of DE-SA is better than that of DE-WF. In fact, the designs 12000 functions in a time between 8 and 10 h. Li et al., (2020)
obtained using DE-SA show heat duty reductions of more employ Simulated annealing for distillation process for se-
than 50% compared with the best heat duty designs obtained parating benzene-isopropanol-water evaluating 81,000 in a
by DE-WF. time of 18 h. Recently, Lyu et al., (2021), report CPU time from
On the other hand, with the aim to provide an alternative 13 to 69 h to evaluate between 40,000 and 140,000 individuals,
visualization of the results, a comparative analysis of DE-SA per run, in the optimization of three separation systems in
and DE-WF was carried out through the box diagram de- ternary mixtures through dynamical DE. They propose the
picted in Fig. 6c). In this scheme it can be observed that the parelization of the evaluation of the objective function re-
median for the DE-SA algorithm, 12.84 GW/y, is less than that ducing substantially the time, however this implementation
for DE-WF, 23.9886 GW/y. Also notice that there is little dis- slightly affects the quality of the solution. They report that
 Chemical Engineering Research and Design 189 (2023) 98–116 113

the complexity of the case study will increase the time it Hence, based on the results obtained, it is clear that the
takes for an Aspen simulation to converge. Some other au- constraint handling technique is a cornerstone for success-
thors omit to report the time and only establish as a point of fully dealing with optimization problems. Besides that, the
comparison the number of function evaluations (Vázquez- experience showed that the dynamic self-adaptive con-
Ojeda et al., 2013). straint handling improves the performance of the DE algo-
rithm. At the same time, it was demonstrated that the
4. Discussion of results implementation of robust statistical tools is an important
factor to support optimization results.
As shown, the optimization performed using DE coupled
with dynamic self-adaptive constraint handling successfully 5. Conclusions
tackled all kinds of optimization problems presented, from
NLP problems to multivariable MINLP problems with high This work introduces a novel self-adaptive constraint hand-
dependence among variables and optimization of a distilla- ling technique coupled to a Differential Evolution algorithm
tion train. The results evidence the robustness of the opti- (DE-SA) based on a self-adaptation dynamic threshold, self-
mization strategy composed of a DE algorithm and dynamic adaptation factors (weights) and added penalization due to
self-adaptive constraint handling supported by a self-adap- an interest in design variables. The optimization perfor-
tation dynamic threshold. In addition, the approach im- mance of DE-SA was compared with that using the con-
plemented demonstrated that the penalization strategy has a straint technique based on a weight function coupled to a
large influence on the trend of the results. This was estab- Differential Evolution algorithm (DE-WF). In addition, to va-
lished as the optimization was conducted using a DE algo- lidate our proposal, the performance of DE-SA was also
rithm in both cases, DE-SA and DE-WF, but the best compared with that of a repair algorithm proposed by
performance was attributed to DE-SA. In general terms, the Kheawhom (2010) to optimize some benchmark functions,
performance of DE-SA shows clear dominance over DE-WF in and with the optimal value determined by Edgar et al. (2001)
terms of the quality of the optimal values obtained (best fit- who used a deterministic method to optimize typical che-
ness function) and numerical effort and computing time. It is mical engineering problems. Finally, a distillation train op-
assumed that this behavior is triggered by means of efficient timization problem was solved with DE-SA and DE-WF. To
exploration of the search space, which is powered by the compare results and perform statistical analysis, we carried
constraint handling technique to an important degree. In out 30 runs under the same conditions, for each technique.
addition, the structure of the penalization, composed of se- The comparative behavior is essentially based on the con-
quential self-adaptive and dynamic penalizations, led the straint handling approaches, not on the optimization algo-
population towards the best values of the objective function. rithm (DE) itself.
Notice that the dynamic self-adaptive constraint handling The optimization results for the benchmark functions
was successfully applied to penalize in a proportional way indicate that the best performance is given by DE-SA. In fact,
according to the degree of deviation in relation to the target the repair algorithm and DE-SA led to the same fitness
established. A key factor in this mechanism is the modula- function optimal values but DE-SA used fewer function
tion of relaxing and stiffening of the dynamic threshold de- evaluations and less computing time than the repair algo-
pending on the behavior of the fitness function of the rithm. In addition, the optimal values obtained by DE-WF
population during the optimization process. were larger than those reported by Kheawhom (2010). Simi-
Thus, through the optimization supported by adequate larly, optimization of the chemical engineering problems
constraint handling (DE-SA), the following benefits were ob- showed that the best fitness function values were obtained
tained: optimal fitness functions were obtained using regular using DE-SA, using less computational time and numerical
numbers of function evaluations, with neither large popula- effort than the optimization process carried out by DE-WF to
tions nor a large number of generations. For instance, com- find the best solution reported by Edgar et al. (2001). In the
pared with a repair algorithm, DE-SA required 20,000 distillation train problem, the best compromise (heat duty
function evaluations, while the repair algorithm performed and total number of stages) for the distillation train was
2000,000 function evaluations. In addition, the optimal in- found using DE-SA. So, designs with both less energy con-
dividual for the repair algorithm was determined in evalua- sumption (heat duty) and a smaller total number of stages
tion 120, whereas the best individual for DE-SA was found in were determined, compared with DE-WF. These conclusions
evaluation 35. On the other hand, compared with DE-WF, a are supported by statistical analysis.
roughly 50% reduction in computing time and better fitness The results shown for DE-SA in all case studies, are at-
functions were determined for DE-SA. In addition, optimi- tributed to the following features:
zation of the distillation train led to a 56% reduction in re-
boiler duty and 26% fewer stages. • Feasible individuals
The optimization results for the distillation train were The feasibility of an individual determines whether that
corroborated by rigorous statistical analysis, using the individual’s information can be used to focus the search in
bootstrap signed-rank test, confidence intervals and box feasible areas and avoid exploration in non-feasible areas.
diagrams. Hypothesis analysis using the bootstrap signed- Thus, no promissory search zones are segregated by the
rank test showed that effectively DE-SA has better perfor- algorithm.
mance that DE-WF. Similarly, the confidence intervals and • Dynamic threshold
box diagrams evidenced that the fitness function obtained The dynamic threshold is applied to inequality constraints
with DE-SA is less dispersed than that computed with DE- so equality constraints must be converted into their in-
WF. In fact, it was obvious that the worst fitness function equality form through a small tolerance value (epsilon).
obtained with DE-SA was better than the best one obtained The main characteristic of the dynamic threshold is the
with DE-WF. relaxation/stiffness of the equality constraints through the
114 Chemical Engineering Research and Design 189 (2023) 98–116

epsilon tolerance, allowing the population to follow the

best search trajectories based on information from pre-
vious generations. Eventually, progressive reduction of the
dynamic threshold enables the population to fulfill all the
equality constraints.
• Self-adaptation factors based on dynamic threshold
Through the self-adaptation of penalty factors based on
the magnitude of constraint violations (according to the
dynamic threshold value), the algorithm identifies and
classifies individuals with the best characteristics and
segregates those whose degree of deviation from a valid
constraint is larger. When the residuals are large, a strong
penalization is applied. For individuals with low con-
straint violation, the penalization is weak.
• Self-adaptation factors (weights) based on number of
violated constraints
This implementation allows identification of the number
of constraints that are violated regardless of their magni-
Fig. A1 – Step 1 of the self-adaptive constraint handling
tude, increasing the effect of the self-adaptation factors
technique.
(see previous paragraph). Thus, for each individual, if the
number of violated constraints is small the penalization is
weak, and it becomes stronger with a larger number of
violated constraints.
• Constraint of design variables

We apply this type of penalty to guide the search towards

solutions to find the appropriate value of a specific design
variable. The proposed penalty method decreases the final
fitness function value only when all constraints lie within the
dynamic threshold, for the minimization problem.
It is evident that the constraint handling technique pro-
posed in this work has an effective selection pressure even
when the total amount of constraint violation decreases.
Additionally, it allows a suitable convergence to the feasible
regions of the search space and leads the population towards
the best values of the fitness function, in all case studies. DE-
SA is a worthy tool since it is able to deal with optimization
problems subject to equality and inequality constraints and
interdependence on the optimization variables. Therefore, to
further evaluate the proposed technique, our future work
aims to couple the dynamic self-adaptive constraint hand-
ling technique to other evolutionary algorithms to optimize Fig. A2 – Step 2 of the self-adaptive constraint handling
more complex chemical engineering processes. technique.

Declaration of Competing Interest

The authors declare that they have no known competing fi-

nancial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Appendix A. Constraint handling with self-

adaptation for the distillation train problem

See. Fig. A1, Fig. A2, Fig. A3. and Fig. A4.
Implementation of the constraint handling technique for
the case of the distillation train is detailed below. The self- Fig. A3 – Step 3 of the self-adaptive constraint handling
adaptive constraint handling approach begins by verifying if technique.
the individual Zi, G is feasible. In this step Q fitness and
violatotal values are return to DE algorithm and shows the
Figure A.1. function, Qmax is assigned. The purities xipur and recoveries
xirec are set to a “bad” value (0.01). In this case, for violapur
• An infeasible individual is when Aspen One non con- and violarec the Ncomponents value are assigned.
vergence (violaconvergenceis1), and worst value at the fitness
 Chemical Engineering Research and Design 189 (2023) 98–116 115

Babu, B.V., Angira, R., 2006. Modified differential evolution (MDE)

for optimization of non-linear chemical processes. Comput.
Chem. Eng. 30 (6–7), 989–1002.
Ben, Hamida, S., 2016. Extension of evolutionary algorithms to
constrained optimization. Metaheuristics. Springer, Cham, pp.
329–356.
Boukouvala, F., Ierapetritou, M.G., 2014. Derivative-free optimi-
zation for expensive constrained problems using a novel ex-
pected improvement objective function. AIChE J. 60 (7),
2462–2474.
Chanthasuwannasin, M., Kottititum, B., Srinophakun, T., 2017. A
mixed coding scheme of a particle swarm optimization and a
hybrid genetic algorithm with sequential quadratic program-
ming for mixed integer nonlinear programming in common
chemical engineering practice. Chem. Eng. Commun. 204 (8),
840–851.
Fig. A4 – Step 4 of the constraint handling approach. Coello, C.A.C., 2002. Theoretical and numerical constraint-
handling techniques used with evolutionary algorithms: a
survey of the state of the art. Comput. Methods Appl. Mech.
Eng. 191 (11–12), 1245–1287.
• For a feasible individual, the fitness function Q fitness is
Coello, Coello, C.A., 1999. Self-adaptive penalties for GA-based
calculated using Aspen One (Qtotal ), violaconvergence is 0, and optimization. Evolutionary Computation (CEC 99). Proc. Congr.
the purities, xipur and recoveries, xirec for each component Evolut. Comput. 1, 573–580.
are reported. Coello Coello, C. A., Mezura Montes, E, 2002. Constraint-
handling in genetic algorithms through the use of dom-
For the feasible individual, we calculate the violation inance-based tournament selection. Adv. Eng. Inform. 16
magnitude, for each component, accordance with dynamic (3), 193–203.
Contreras - Zarazúa, G., Jasso - Villegas, M.E., Ramírez - Márquez,
threshold value, . The magnitude are obtained with the re-
pur C., Sánchez - Ramírez, E., Vázquez - Castillo, J.A., Segovia -
sidual between the target value x target and the achieved value
Hernández, J.G., 2021. Design and intensification of distillation
xipur of purities. With residual, wipur and Qpur is calculated. processes for furfural and Co-products purification con-
For recoveries likewise is applied and obtain the wirec and Qrec . sidering economic, environmental, safety and control issues.
Finally, computes the Qpen adding Qpur , Qrec and Qtotal . When Chem. Eng. Process. Process.Intensif. 159, 108218.
the individual meet whole constraints, Qpen is not pena- Dadios, E., Ashraf, J., 2006. Genetic algorithm with adaptive and
lized and assigned at Qnum_pen (see Figure A.2). In addition, we dynamic penalty functions for the selection of cleaner pro-
duction measures: a constrained optimization problem.
determine the total number of no accomplished con-
Clean. Technol. Environ. Policy 8 (2), 85–95.
straints violatotal with the sum of violapur and violarec values
Dasgupta, D., Michalewicz, Z., 1997. Evolutionary algorithms—an
(see Figure A.3). These values are obtained count the viola- overview. Evolutionary Algorithms in Engineering
tions of purities and recoveries for each component for de- Applications. Springer, Berlin, Heidelberg, pp. 3–28.
termines, wnum_p en and Qnum_pen . In this step Qnum_pen and Deb, K., Agrawal, S., Pratap, A., Meyarivan, T., 2000. A fast elitist
violatotal values are return to DE algorithm. non-dominated sorting genetic algorithm for multi-objective
Finally, only individuals that accomplish whole con- optimization: NSGA-II. In International conference on parallel
problem solving from nature 849–858. https://doi.org/10.1007/
straints decrease the fitness function. This penalization favor
3-540-45356-3_83
the search in feasible spaces where the heat duty is mini- Diwekar, U.M., Grossmann, I.E., Rubin, E.S., 1992. An MINLP pro-
mized and also reduces the total number of stages of the cess synthesizer for a sequential modular simulator. Ind. Eng.
distillation scheme. We determines the factor wstages with the Chem. Res. 31 (1), 313–322.
relationship between a design value heuristic, NTmax and total Edgar, T.F., Himmelblau, D.M., Lasdon, L.S., 2001. Optimization of
number of stages, NTtotal . When NTtotal < NTmax ob- Chemical Processes. McGraw-Hill, Boston.
tain Q pen_stages . Fister, I., Brest, J., Iglesias, A., Galvez, A., Deb, S., 2021. On selec-
tion of a benchmark by determining the Algorithms’ qualities.
that and assign to Q fitness, otherwise Q fitness is Q num_pen (see
IEEE Access 9, 51166–51178. https://doi.org/10.1109/ACCESS.
Figure A.4).
2021.3058285
Franke, M.B., 2019. Mixed-integer optimization of distillation se-
quences with Aspen Plus: a practical approach. Comput.
Chem. Eng. 131, 106583.
Gómez – Castro, F.I., Segovia - Hernández, J.G., Hernández, S.,
References Gutiérrez – Antonio, C., Briones – Ramírez, A., Gamiño –
Arroyo, Z, 2015. Design of non-equilibrium stage separation
systems by a stochastic optimization approach for a class of
Adeli, H., Cheng, N.T., 1994. Augmented Lagrangian genetic al-
mixtures. Chem. Eng. Process.: Process.Intensif. 88, 58–69.
gorithm for structural optimization. J. Aerosp. Eng. 7 (1),
Hesterberg, T., 2011. Bootstrap. Wiley Interdiscip. Rev.: Comput.
104–118.
Stat. 3 (6), 497–526.
Asafuddoula, M., Ray, T., Sarker, R., 2015. A differential evolution
Homaifar, A., Qi, C.X., Lai, S.H., 1994. Constrained optimization
algorithm with constraint sequencing: An efficient approach
via genetic algorithms. Simulation 62 (4), 242–253.
for problems with inequality constraints. Appl. Soft Comput.
Joines, J., Houck, C., 1994. On the use of non-stationary penalty
36, 101–113.
functions to solve nonlinear constrained optimization pro-
Aspen Plus 13.0, 2007. User Models. Aspen Technology, Inc.
blems with GA's. Proc. 1st IEEE Conf. Evolut. Comput., Orlando
Austbø, B., Gundersen, T., 2016. Impact of problem formulation
579–584.
on LNG process optimization. AIChE J. 62 (10), 3598–3610.
Kazarlis, S., Petridis, V., 1998. Varying fitness functions in genetic
Austbø, B., Wahl, P.E., Gundersen, T., 2013. Constraint handling in
algorithms: Studying the rate of increase of the dynamic
stochastic optimization algorithms for natural gas liquefac-
penalty terms. International Conference on Parallel Problem
tion processes. Computer Aided Chem. Eng. 32, 445–450.
116 Chemical Engineering Research and Design 189 (2023) 98–116

Solving from Nature. Springer,, Berlin, Heidelberg, pp. Srinivas, M., Rangaiah, G.P., 2007. Differential evolution with tabu
211–220. list for solving nonlinear and mixed-integer nonlinear pro-
Kheawhom, S., 2010. Efficient constraint handling scheme for dif- gramming problems. Ind. Eng. Chem. Res. 46 (22), 7126–7135.
ferential evolutionary algorithm in solving chemical engineering Storn, R.M., Price, K.V., 1997. Differential evolutions a simple and
optimization problem. J. Ind. Eng. Chem. 16 (4), 620–628. efficient heuristics for global optimization. J. Glob. Optim. 11,
Kiss, A.A., Segovia-Hernández, J.G., Bildea, C.S., Miranda-Galindo, 341–359.
E.Y., Hernández, S., 2012. Reactive DWC leading the way to Summanwar, V.S., Jayaraman, V.K., Kulkarni, B.D., Kusumakar,
FAME and fortune. Fuel 95, 352–359 In press. H.S., Gupta, K., Rajesh, J., 2002. Solution of constrained opti-
Leboreiro, J., Acevedo, J., 2004. Processes synthesis and design of mization problems by multi-objective genetic algorithm.
distillation sequences using modular simulators: a genetic Comput. Chem. Eng. 26 (10), 1481–1492.
algorithm framework. Comput. Chem. Eng. 28 (8), 1223–1236. Takahama, T., Sakai, S., 2006. Constrained optimization by the ε
Liepins, G., Vose, M., 1990. Representational issues in genetic constrained differential evolution with gradient-based muta-
optimization. J. Exp. Theor. Artif. Intell. 2 (2), 101–115. tion and feasible elites. IEEE Int. Conf. Evolut. Comput.
Lyu, H., Cui, C., Zhang, X., Sun, J., 2021. Population-distributed Vanc. 1–8.
stochastic optimization for distillation processes: im- Takahama, T., Sakai, S., 2010. Constrained optimization by the ε
plementation and distribution strategy. Chem. Eng. Res. Des. constrained differential evolution with an archive and gra-
168, 357–368. dient-based mutation. IEEE Congr. Evolut. Comput. 1–9.
Lyu, H., Zhang, X., Cui, C., Sun, J., 2022. Adaptive superstructure Teh, Y.S., Rangaiah, G.P., 2003. Tabu search for global optimiza-
for multiple-interconnection process synthesis: Eliminate tion of continuous functions with application to phase equi-
unnecessary flowsheet predetermination to reduce com- librium calculations. Comput. Chem. Eng. 27 (11), 1665–1679.
plexity. Chem. Eng. Process. -Process. Intensif. 171, 108731. Vazquez–Castillo, J.A., Venegas–Sánchez, J.A.,
Medina - Herrera, N., Tuttuti - Ávila, S., Jiménez - Gutiérrez, A., Segovia–Hernández, J.G., Hernández-Escoto, H., Hernandez,
Segovia - Hernández, J.G., 2017. Optimal design of a multi- S., Gutiérrez–Antonio, C., Briones–Ramírez, A., 2009. Design
product reactive distillation system for silanes production. and optimization, using genetic algorithms, of intensified
Comput. Chem. Eng. 105, 132–141. distillation systems for a class of quaternary mixtures.
Mezura Montes, E., Coello Coello, C. A, 2011. Constraint-handling Comput. Chem. Eng. 33 (11), 1841–1850.
in nature-inspired numerical optimization: past, present and Vázquez-Ojeda, M., Segovia - Hernández, J.G., Hernández, S.,
future. Swarm Evolut. Comput. 1 (4), 173–194. Hernández – Aguirre, A., Kiss, A.A, 2013. Design and optimi-
Michalewicz, Z., Attia, N., 1994. Evolutionary optimization of zation of an ethanol dehydration process using stochastic
constrained problems. Proc. 3rd Annu. Conf. Evolut. Program. methods. Sep. Purif. Technol. 105, 90–97.
Singap. 98–108. Wong, K.P., Dong, Z.Y., 2005. Differential evolution, an alternative
More, R.K., Bulasara, V.K., Uppaluri, R., Banjara, V.R., 2010. approach to evolutionary algorithm. Proc. 13th Int. Conf.
Optimization of crude distillation system using aspen plus: Intell. Syst. Appl. Power Syst. 73–83.
Effect of binary feed selection on grass-root design. Chem. Yiqing, L., Xigang, Y., Yongjian, L., 2007. An improved PSO algo-
Eng. Res. Des. 88 (2), 121–134. rithm for solving non-convex NLP/MINLP problems with
Na, J., Lim, Y., Han, C., 2017. A modified DIRECT algorithm for equality constraints. Comput. Chem. Eng. 31 (3), 153–162.
hidden constraints in an LNG process optimization. Energy Yu, X., Lu, Y., Wang, X., Luo, X., Cai, M., 2019. An effective im-
126, 488–500. proved differential evolution algorithm to solve constrained
Palma - Barrera, J.P., Sánchez - Ramírez, E., Ramírez - Márquez, optimization problems. Soft Comput. 23 (7), 2409–2427.
C., Cervantes - Jauregui, J.A., Segovia - Hernández, J.G., 2019. Zahara, E., Kao, Y.T., 2009. Hybrid Nelder–Mead simplex search
Reactive distillation column design for tetraethoxysilane and particle swarm optimization for constrained engineering
(TEOS) production. Part II: dynamic properties and inherent design problems. Expert Syst. Appl. 36 (2), 3880–3886.
safety. Ind. Eng. Chem. Res. 58, 259–275. Zhang, H., Rangaiah, G.P., 2012. An efficient constraint hand-
Riche, R.L., Haftka, R.T., 1997. Evolutionary optimization of ling method with integrated differential evolution for nu-
composite structures. Evolutionary Algorithms in Engineering merical and engineering optimization. Comput. Chem. Eng.
Applications. Springer, Berlin, Heidelberg, pp. 87–102. 37, 74–88.
Sreepathi, B.K., Rangaiah, G.P., 2017. Optimization of heat ex- Zielinski, K., Laur, R., 2008. Stopping criteria for differential evo-
changer network retrofitting: comparison of penalty function lution in constrained single-objective optimization. Advances
and feasibility approach for handling constraints. Multi in Differential Evolution. Springer, Berlin, Heidelberg, pp.
Object. Optim. Tech. Appl. Chem. Eng. 501–532. 111–138.