Grey wolf optimizer trained ANN technique for development of explainable model of commercial ethylene oxide reactor and multi- objective optimization to maximize profit

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 11 | Nov 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 404
Grey wolf optimizer trained ANN technique for development of
explainable model of commercial ethylene oxide reactor and multi-
objective optimization to maximize profit
Abstract - The current research focuses on the use of a grey wolf optimizer trained ANN for thebuildingofanindustrialreactor
model, as well as its optimization for profit maximisation. As a case study, an ethylene oxide(EO)reactorwasused. TheEO reactor
is a multiphase catalytic reactor with a complex nature that makes developing a reliable first principle-based model challenging,
while data-driven models lack explanability. As a result, a general methodology has been devised in this study, in which a data-
driven technique such as grey wolf optimizer trained ANN was utilised as a modelling tool, and themodel wasthenpost-processed
to improve the model's explanability. The performance of the model in capturing the system's inherent physics was used to make
the decision. Following the creation of an appropriate model, it was subjected to optimization. Maximizationofcatalystselectivity
and minimization of reactor temperature were examined here, both of which have a significant impactontheplant'sprofitability.
Because the two aims are incompatible, a pareto solution was found by using a multi-objective evolutionary algorithm on the
generated model.
Key Words: Grey wolf optimizer trained ANN; Ethylene oxide reactor; Modelling; Optimization; Genetic algorithm
1.INTRODUCTION
Due to intense global competition, the business environment has changed dramatically since last decade. Chemical process
industries are seeing their profit margins dwindle as a result of increased cutthroat competition in a volatilemarket,thanksto
globalization. Furthermore, the developing environmental concernof global warming haspromptedgovernmentauthoritiesin
several countries to establish rigorous environmental regulations, whichaddtothe expenseofdoingbusiness.Theonlywayto
successfully minimize these difficulties is through technological innovation to improve process and energy efficiency, yield
maximization, and environmental impact reduction. Chemical firms all around the world are looking for new and imaginative
methods to save costs and increase profits. One of the most appealing creative strategies to explore is the use of artificial
intelligence-based techniques to extract value from massive amounts of plant operating data through data mining and
knowledge discovery.
Chemical reactors are attracting the attention of researchers and plant managers who arelookingfornew waystocreatemore
profit. The reactor is the only significant piece of equipment that adds real value to raw materials by converting them into
finished products. Reactor optimization has a huge potential impact on total plant profitability in this regard [1]. Modeling
complicated systems of industrial chemical processes is essential as a first step in optimizingthecommercial reactor.Complex
reaction kinetics and thermodynamics were involved in industrial chemical reactions. Building a reliable phenomenological
model for commercial reactors is a time-consuming and challenging task that necessitates a thorough understanding of
industrial heterogeneous catalytic behavior such as mass diffusion, catalyst deactivation, and so on. Furthermore, several
parameters such as agitation speed, catalyst interfacial area, negative influence of poisons present inincominggas,diffusional
coefficients, and others affect reaction rate, yield, and selectivity in commercial plants, the mechanisms of which are poorly
understood. As a result, a lack of understanding of chemical reaction kineticsobstructsthereactor'soptimization. Thereactors
in most chemical plants remain a black box, and operation engineers do not tamper with them in a live plant for safety and
reliability reasons. This leads to non-optimal operating of industrial reactors, which has a significant impact on the plant's
profitability. As a result, reactors are regarded as uncharted area inthechemical industry.Inlarge-scaleoperations, however,a
slight increase in catalyst selectivity and reaction yield has a significant impact on raw material consumption and plant
profitability [1].
Typically, laboratory scale kinetic analysis is performed in a very ideal and controlled settingtoderivethekinetic equationofa
particular commercial reaction. Because of the vastscaleofindustrial reactors,theexistenceofpoisonsandinertin input gases,
and the varied heat and mass transfer environment, the use of such a method for kinetic modelling of industrial reactors is
debatable. Furthermore, due to the limited market window for chemical goods, plant owners are hesitant to devote sufficient
time and effort to a thorough examination of these complex reaction mechanisms. As a result, an alternative simple way is to
1,2
2
Sucharita Pal1 , Sandip K Lahiri ,
Chemical Engineering Department, National Institute of Technology Durgapur, Durgapur – 713209
---------------------------------------------------------------------***---------------------------------------------------------------------

use a data-driven effective computational strategy to construct approximated reactor models for these complex reactions
systems which can be subsequently used to optimize the reactor and increase profit.
Because most chemical plants gather and store massiveamountsof reactorintakeand exithistoricplantoperational data every
minute, the true challenge is figuring out how to use this wealth of data to earn more money. Data is the new oil, and data-
driven modelling approaches like artificial neural networks (ANN)andsupportvector machines(SVM)areunquestionablythe
new IC engines of our time. ANN and SVR have been highly prominent black box modelling approaches in the recent decade,
with a number of applications developed for commercial reactors. Plant engineers, on the other hand, dislike ANN and SVM
models because of their lack of explainability and the model's black box nature. ANN model does not provide any insights of
underlying physics of chemical reactors.
Because ANN modelling gives a black box type equation consistingofa complexsigmoidal functionwithmultipletuningfactors
known as weights and biases, the model suffers from explainability limitations, despite its great prediction capacity. Process
engineers typically desire comprehensible equations in differential/ algebraic form that relate output variables to input
process parameters so they can gain more insight and benefit. It is preferable that the generated model include a closed form
equation that can describe the effect of key process parameters on the output variable. SVM has the same problem as ANN in
that none of them provides closed form explainable equations that are portable and easy to implement in a plant DCS system.
For plant engineers, gaining insights and obtaining a closed form explainable model equation is critical to their acceptance of
using the model in the plant.
Despite the ANN approach's impressivepredictioncapabilities,applicationsofthistechnologytochemical reaction engineering
jobs in commercial reactors are conspicuously lacking. In this study, an attempt was made to apply ANN modelling to a
commercial chemical ethylene oxide (EO) reactor. One of the key goals of this research is to turn the ANN model into an
explainable closed form equation, which will reveal important information about the phenomenology of the reactor.
The second goal of this work is to use recently discovered nature-inspired grey wolf optimizationapproachestotraintheANN
model. Learning is an important component of every neural network, and it is attracting the attention of many academics.The
traditional [7] or enhanced [8– 10] Back-Propagation (BP) algorithm has been used to train feedforward neural networks
(FNNs) in most applications. The BP method is a gradient-based algorithm with some flaws, including delayed convergence
[11] and a possibility to get stuck in local minima [12]. During the learning process of FNNs, the goal is to identify the best
combination of connection weights and biases to achieve the least amount of error. FNNs, on the other hand, frequently
converge on points that are the best answer locally but not globally. In other words, rather than the global minimum, learning
methods lead FNNs to local minima. According to [11], the BP algorithm's convergence is strongly dependent on the weights,
biases, and parameters' initial values. Using unique heuristic optimization methods or evolutionaryalgorithmstoimprovethe
problems of BP-based learning algorithms is a common option in the literature.
Simulated Annealing (SA) [13,14], Genetic Algorithms (GAs) [15], Particle Swarm Optimization (PSO) algorithms [16–20],
Magnetic Optimization Algorithm (MOA) [21], and Differential Evolution (DE) [22] are some of the heuristic optimization
methods that have been utilized to train FNNs. According to [11], some algorithms,suchasSAandGA,canlessenthelikelihood
of local minima trapping, although they still have poor convergence rates.
Despite the fact that numerous meta-heuristics have been used widely in ANN learning, none ofthemhasperformed well inall
applications. Furthermore, existing metaheuristics have a number of drawbacks [23-27], including slow convergence speed,
trapping in local minima, long computational time, tuning many parameters, and a difficult encoding scheme. As a result, it
appears that improving the performance of prior meta-heuristics or proposing new ones is required to improve the efficiency
of ANN learning in various domains. GWO is a recent innovative stochastic and metaheuristic optimization approach,whichis
introduced by [18]. In this paper, the efficiencies of GWO technique are investigated for training FNNs.
The current study opted to model a commercial EO reactor because of its significant financial significance in worldwide
petrochemical markets. Because it is used to make polyester(whichisusedtomaketextile)andPETbottles,theethyleneglycol
industry is regarded the second most significant petrochemical product in terms of volume, profit margins, and global market
demands. Recently, the glycol sector has seen intense competition in the global market, with players vying for higher reactor
yield and selectivity, which has a significant impact on production costs.
The next goal of this study is to use the developed model to raise the profit of the ethylene oxide plant while reducing its
environmental impact. This is accomplished by using a model-based, nature-inspired metaheuristic optimization method to
optimize the input process parameters in order to improve selectivity (i.e., reactor performance) while lowering reactor
temperature (which maximize catalyst longevity). The two main operational objectives in commercial EO reactors, namely,

maximize selectivity (i.e., maximize profit) and minimize reactor temperature (corresponds to greater catalystlongevity),are
mutually exclusive and conflicting, and one cannot achieve the highest selectivity without sacrificing temperature, and vice
versa. The multi objective genetic algorithm (GA) is used to improve the input space of the EO reactor ANN model in order to
provide pareto optimal solutions that meet both objectives in the most efficient way feasible.
2. Case study of ethylene oxide reactor
2.1 Background
Because of the lack of understanding and complexity of multiphase catalytic processes in commercial ethylene oxide reactors,
trustworthy first principle-based modelsare hard to come by. Data driven modelling isa feasible alternativestrategyduetothe
vast amount of reactor operating parameter data available from commercial reactors. The current projectaimstomakeuseofa
huge amount of process data to create a framework for converting the data's information to knowledge and subsequently into
profit.
2.2 Reactions:
The manufactureof ethylene oxide (EO) is thefirststepintheethyleneglycolmanufacturingprocess.Ethyleneoxideisproduced
in commercial glycol plants by reacting ethylene with oxygen (Equation 1) at 230–270°C and 20 bar pressures. In addition to
ethylene oxide, an unfavorable sidereaction (Equation 2) produces carbon di-oxide, which is released into the atmosphereand
has a severe negative influence on the environment.Catalystselectivityiscomputedusingequation3andmeasurestheextentof
the first reaction (desired) relative to the second reaction (undesired). Selectivity maximization is a key production goal for
increasing profits while minimizing environmental damage.
(1)
(3)
2.3 Process flow diagram
Figure1 [17] shows a simplified process flow diagram forthe synthesis of EO. The EOreactorisashellandtubereactorinwhich
the catalyst is put in the tube side as packed bed and the reactions described in Equations 1 and 2 occur in this part. Water is
cycled through the shell sideto recover heat because the reactionsare exothermic. Because the conversion of EO fromethylene
and oxygen is very low par pass, the unreacted reactants are recycled back through the cycle gas compressor to boost the
conversion. In industrial, unreactedethylene,unreactedoxygen,carbondioxide,methane,EO,nitrogen,(usuallycalledcyclegas)
are combined with fresh ethylene and oxygen streams. The reactor output gas was routed through EO absorbers and CO2
absorbers, as shown in Figure 1. The EO is selectively absorbed by water in an EO absorber, whilst the CO2 is absorbed by a
carbonate solution in a CO2 absorber. Plant's overall profitability is determined by the catalyst's selectivity, and plant aims to
decreases the second undesired reaction while maximizing the first favorable reaction.
(2)

FIGURE 1: Process flow diagram of EO production process
2.4 Production Objectives:
The catalyst selectivity declines gradually with the age of the catalyst due to catalyst poisoning and sintering effects, and the
catalystselectivity maintains about 90% at the start of operation (SOR) and dropsto80%attheendofoperation(EOR)(usually
after 2-4 years). The average catalytic life of today's catalysts is between 24 and 48 months. The temperature of the reactor is
gradually increased over the lifeof the catalyst [say from 225 ˚C (SOR) to 275˚C(EOR)]toenhancethereactionrateandkeepthe
plant production on target while selectivity is dropping. The selectivity of the catalyst impacts the plant's profitability, asbetter
selectivity means less consumption of expensive ethylene and oxygen to produce same amount of EO i.e., less production cost.
Reactor temperature, on the other hand, determines catalyst longevity. A faster rise in reactor temperature causes a loss of
catalystage and a higher fixed cost ofcatalyst replacement, which necessitatesa plant downtime. Optimizing reactoroperating
parameters in such a way that selectivity maximization and reactor temperature minimization (i.e., maximization of catalyst
longevity) are achieved simultaneously while maintaining plant production targets are two primary production objectives in
commercial plants.
3. Hybrid FNN- GWO methodology
This section explains the basic background materials neededforathoroughcomprehensionofthesuggestedmethod.TheMulti-
layer perceptron (MLP) as a Feed-forward neural network (FNN) is first presented, followed by the GWO methods, which are
then compared to the suggested technique for MLP learning in industrial reactors modelling. A quick explanation of the hybrid
ANN and GWO training process follows MLP network.
Because of its capability to map any function to an indefinitedegree of precision, the feed-forward neural network(FNN)isone
of the most prevalent ANNs used to solve real lifeproblems and attracts a lot of academic attention. The multi-layer perceptron
has been extensively used in finance, medicine, engineering,geology,physics,andbiology,amongotherfields.Ithasbeenwidely
used in the chemical industry for nonlinear process modelling, fault diagnostics, and process control.
MLP has one input layer that receives external inputs, one or more hidden layers, and one output layer that displaystheoutput,
as shown in Figure 2. Except for the input layer, all levels are built up of processing nodes and activation functions. Data is
supplied at the input layer, and network nodes perform calculations in successive layers until each output node obtains an
output value.
3.1 Training of ANN by GWO
In general, there are three ways to train FNNs with a heuristic algorithm. First,heuristictechniquesareusedtodiscoverthebest
combination of weights and biases for a FNN with the least amount of error. Second, heuristic algorithms are usedtodetermine
the best structure for a FNN in each problem. The final way is to tune the parameters of a gradient-based learning system, such
as the learning rate and momentum, using an evolutionary algorithm.

Before training FNNs, the structure is fixed in the first scenario. A training algorithm's job is to discover a good value for all
connection weights and biases in order to reduce the overall error of the FNNs.FNN structures differ in the second scenario.To
discover the appropriate structure for a given problem, a training procedure is applied to an FNN. Manipulate the connections
between neurons, the number of hidden layers,and the numberofhiddennodesineachlayeroftheFNNtochangethestructure.
GWO is applied to a FNN using the first technique in this paper; these processesare referred toashybridFNN-GWO.Thismeans
that the FNN'sstructureis fixed, and the GWO algorithm finds a set of weightsandbiasesthatyieldstheleastamountoferrorfor
the FNN. The following essential elements must be defined in order to design FNN-GWO. To evaluate fitness in FNN-GWO, first
develop a fitness function based on the FNN's error. Second, for the agents of FNN-GWO, an encoding approach should be
established to encode the FNN's weights and biases. These components are detailed in further detail below.
3.2 Fitness function
The following is the formula for the fitness function used in this article [10]:
The number of input nodes is equal to n, the number of hidden nodes is equal to h, and the number of output nodes isequaltom
in Figure 2, which has two levels (one input, one hidden, and one output layer). Each hidden node's output is calculated as
follows at the end of each learning epoch:
, j = 1, 2, …, h (4)
In , n is the number of the input nodes, wij is the connectionweight from the ith nodein theinputlayertothe
jth node in the hidden layer, Ɵj is the bias (threshold) of the jth hidden node, and xi is the ith input.
After calculating outputs of the hidden nodes, the final output can be defined as follows:
, k = 1,2, …, m, (5)
where wkj is the connection weight from the jth hidden nodeto the kth output nodeandƟk isthebias(threshold)ofthekth output
node.
Finally, the learning error E (fitness function) is calculated as follows:
(6)
(7)
where q is the number of training samples, is the desired output of the ith input unit when the kth trainingsampleisused,and
is the actual output of the ith input unit when the kth training sample is used. Therefore, the fitness function of the ith training
sample can be defined as follows:
Fitness (Xi) = E(Xi)

FIGURE 2: FNN with a 2-3-1 structure
3.3 Encoding strategy
Following the specification ofthe fitness function for FNN-GWO, the next step is to chooseanencodingstrategyforeachagentin
FNN-GWO to encodethe FNN's weightsand biases. In matrix encodingstrategy,eachagentisencodedasamatrix.Anexampleof
this encoding strategy for the FNN of Figure 2 is provided as follows:
, (8)
, , , , (9)
where W1 is the hidden layer weight matrix, B1 is the hidden layer bias matrix, W2 is the output layer weight matrix, is the
transpose of W2, and B2 is the hidden layer bias matrix.
3.4 Grey Wolf Optimization (GWO): at a glance
Mirjalili et. Al. [28] introduced GWO, a new stochastic andmetaheuristic optimizationalgorithm.Twoimportantcharacteristics
of meta heuristic optimization are exploration and exploitation. Exploration means how quicklyanoptimizationalgorithmable
to identify the potential area where the global minima may exist from all feasible large search area. Once this potential area is
found, exploitation means how quickly and accurately detect the exact point where the global minima exist by detail search. In
literature, it is found that GWO algorithmable to make a delicatebalancebetweenexplorationandexploitationbyemulating the
cooperative hunting behavior of grey wolves in nature. This bionic optimization method promotes the grey wolf pack's
hierarchical system and predation behavior, in which the wolves take prey by surrounding, haunting,andattackingitunderthe
command of the head grey wolf [31]. In this large-scale search methodology, all wolves form a rigid hierarchical pyramid as
depicted in Figure 3. but decision is based on the top three grey wolves siting at the top of pyramid.

FIGURE 3: Hierarchy of grey wolf (dominance decreases from top down)
The average group size is 5-12 wolves. The top layer (α layer), which consists of a male and female leader, is the team's most
powerful and capable member for makingdecisionsconcerningpredationandotheractivities.Thesecondandthirdlayersinthe
hierarchy, β and δ layer respectively, are responsible for assisting in the behavior of group organizations. Most of the overall
populace resides in the bottom of the pyramid, which is referred to as ω layer. They are primarily in charge of satisfying the
entire pack by balancing the population's internal relationships, caring for the young, and sustaining the dominance hierarchy
[28].
3.5. Mathematical modelling of GWO:
The main key point of the GWO model is the social hierarchy, encircling, hunting, attacking and searching prey.
A) Social hierarchy
In the model of GWO, α is considered as fittest solution. β and δ are considered as second and third best solution, respectively.
The rest of the solutions are assumed to be ω.
B) Encircling
At first, the location of the prey is determined and during the hunting process grey wolves encircled the prey. The following
equations are proposed for mathematically modelling [28].
(10)
(11)
Where is the number of current iterations, is the position vector of one grey wolf, is the next position vector it
arrives, is the position vector of the prey, and are coefficient vectors which are evaluated as follows:
(12)

(13)
and are random vectors in [0,1], a is decreasing value during the iteration in [0,2], typically (I is the
maximum number of iterations).
In thisconcept, grey wolves move around the best solution in hyper-cubes withinan n dimensional space andable to detectthe
position of the prey and encircle it.
C) Hunting
After encircling the prey, grey wolves have the ability to hunt the prey with the guidance of alpha. The beta and delta also take
part occasionally in hunting process. In the mathematical stimulation of hunting behaviour, thefirstthreebestsolutionsupdate
the position of other search agents (including the omegas). The following equations are proposed in this regard.
(14)
(15)
(16)
D) Attacking prey (exploitation)
Grey wolves complete the hunting phase and ready to capturethe prey. For the purposeof mathematically modelling,thevalue
of gradually decrease. Therefore, the fluctuation rateof isalsodecreasedby . whereaisdecrementedfrom2
to 0 over the course of iterations. When the random valuesof , the next position of searchagentcanbeinanyposition
between its current positionand the position of the prey.When ,thegreywolveswoulddivergefromthepreytoachieve
global search and when , the grey wolves would converge towards the prey and complete it.
E) Searching prey (exploration)
GWO algorithm uses an efficient exploration methodology by allowing its search agents to update their position on the basis of
alpha, beta, delta and attacks towards the prey. This mechanism creates a good diversity in the problem search space. Grey
wolves are stay away from each other for global search of prey and close to each other for attacking the prey. For , values
between 1 and -1 are taken. vector is also favoured the exploration technique and random values are used. After
generating the random population, alpha, beta and delta determined the position of the best prey. For selection of exploration
and exploitation, the value of is decreased from 2 to 0 respectively.
The GWO algorithm terminates when the criterion is satisfied. This metaheuristic approach is applied in various real-world
problems because of its efficient and simple performance ability by tuning the fewest operators [28-32].

FIGURE 4: Pseudo code of Grey Wolf Optimisation
3.6. Hybrid learning FNN network by GWO
Essentially the training algorithm of FNN is an optimization problem where optimized set of weigths and biases are to be
determined which minimizesthe error (MSE) beweenactualoutputandpredictedoutput.TheoptimizationcapabilityofGWOis
utilised for FNN learning in this part. To improve the accuracy of the network, a hybrid learning of GWO and FNN networks
(GWO-FNN) is used. By training the network, the algorithm will concurrently determine the set of weights and their
corresponding accuracy. The network weights and biases of the FNN network can be represented as a D-dimensional vector.
Equation 17 defines the vector for FNN (4). The dimension of each particle is treated as the vector D as given below:
D = (Input × Hidden) + (Hidden × Output) +Hidden bias + Output bias, (17)
where Input, Hidden and Output are referred the number of input, hidden and output neurons of FNN network respectively.
The number of biases in the hidden and output layers is also known as Hidden bias and Output bias. Figure 5 depicts the
flowchart of GWO and FNN hybrid leaning ( GWO-FNN). A dataset is collected, normalised, and read to begin the GWO-FNN.
Following that, the appropriate number of input, output,and hidden neuronsarespecifiedtoestablishtheparticledimensionas
Equation (7). The population is initialised, and the training error is determined as a fitness function given in equation 17
followingFNN training. Every particle (wolf) modifiesitsvelocityandpositionbasedontrainingerror.Thenewplacesrepresent
the FNN network's new weights, which are supposed to minimise thefitnessfunction.Thefitnessfunctioniscomputedbasedon
test set error. These steps will go on until meeting stop conditions.

FIGURE 5: Flowchart of Hybrid FNN-GWO methodology

4. Mathematical modelling and optimization of EO reaction process
4.1 Selection of input and output variables for modelling
Because selectivity and reactor temperature have such a significant impact on plant profitability, they are preserved as output
variables. Selectivity was chosen because it directly depicts plant production costs and has a significant impact on plant
profitability. Reactor temperature, on the other hand, impacts catalyst longevity and thus adds a fixed cost to plant production.
Faster temperature rises imply a shorter catalyst life and a higher catalyst cost. As a "wish list" of input variables, all reactor
operational parameters that potentially effect selectivity and reactor temperature are kept. For the first three years, all EO
reactor operational data (daily average) was gathered. Then, after consulting with a plant expert, all of the input variables that
could affect the output variables were recorded. Following that, a cross-correlation study was performed. The correlation
coefficients of each of the input variables with each of the output variables, as well as the inter input cross-correlation
coefficients, were determined using this procedure.
The input variables are shortlisted using the criteria below.
(i) There should be a strong cross-correlation coefficient between input and output variables for a given input variable.
(ii) The values of inter input variable cross-correlation coefficients should be low.
(iii) To keep the model as simple as feasible, the input set of variables was kept to a minimum.
Eight input variables are finally shortlisted and tabulated in Table 1 based on the aforesaid criteria.
TABLE 1 Input Output variables for model building and their range
Variables used in modeling Data Range
Input Variables
Oxygen inlet concentration, mole% (x1) 4.38 - 7.75
C2H4 inlet concentration, mole% (x2) 17.69 - 34.97
CO2 inlet concentration, mole% (x3) 0.21 - 0.45
CG Pressure, bar (x4) 21.84 – 22.64
CG Flow (MT/h) (x5) 687.78 – 803.39
Work Rate (Kg/hr. /m3) (x6) 104.38 – 202.47
Cumulative EOE (MT/m3) (x7) 275.28 – 4469.32
Total chloride concentration, PPM (x8) 2.51 – 5.86
Output Variables
Catalyst selectivity, % (Y1) 81.00 – 89.44
Reactor Temperature, °C, (Y2) 225.83– 250.75
4.2. Data collection, Data cleaning and removal of outliers
Data historian software is used in commercial facilities to gather and store every operating parameter at every minute. We
acquired EO reactor operating data (daily average) for all parameters listed in Table 1 from a commercial plant that had been
operational for nearly three years. Following that, the raw data is subjected to data cleaning and outlier removal tools.
Because industry data contains noise, spikes, outliers, and error as a result of malfunctioning sensors, transmitters, analyzers,
control systems, and data historian software, data cleansing isoneofthemostimportant tasksinindustrialprocessmodelling.It
is commonly established that the quality of data utilized to construct a data-driven model determines the model's quality. Data
quality is an important issue to consider when using data driven modelling because noisy and faulty data can have a significant

impact on model performance. Due to the large volume of process data, this study created an automatic effective data cleaning
technique to prevent manual cleaning. From the plant operating dataset, an automated MATLAB-based algorithm was
constructed to generate a multivariate statistical vector called t-squared. The corresponding rows in the t-squared vector with
values over the 95th percentile was therefore considered outliers and were eliminated from the dataset.
4.3. Modelling through FNN-GWO algorithm
The cleansed data was used to create an ANN-based model. The dataset, which included eight input variables and two output
variables (Table 1), was randomly partitioned into a trainingset(80%ofthetotaldata)and atestset(20%ofthetotaldata).The
training data set was used to compute the model by maximizing the fitness value, while the test data set was utilized to cross-
validate the produced expression. The basic goal of cross-validation is to improve the model's generalizability.
The FNN-GWO-based model was created using code written in MATLAB 2019a. The fitness function used in this work was the
mean-squared error (MSE) between actualand predictedoutputs,and the program was run in such a waythattheMSEvalueof
test set data was minimized. The software was run 100 times to construct the model due to the stochastic nature of the ANN.
4.4. Optimization through Multi objective grey wolf optimization (MOGWO)Algorithm
The model is then used to optimize the reactor process parameter once a reliableand accurateEO reactor modelhasbeenbuilt.
The goal is to establish the ideal process condition that maximizes the profitability of the EO plant, and to do so, the following
conflicting criteria must be met:
(i) The selectivity of the catalyst must be maximized.
(ii) The reactor temperature must be kept as low as possible.
In the present work, multi-objective grey wolf optimization (MOGWO) algorithm is used to strike a balance between two
conflicting objectives. MOGWO has proven itself to be an excellent multi objective optimization tool and has been used
extensively in diverse field of engineering and medicine.
As there are two conflicting objectives in this particular case study, MOGWO is used to generate pareto optimal solutions. For
implementation of MOGWO algorithm, two objective functions were developed which are as follows (Equation 18 and 19):
F1(x) = 1/(1+Y1(x)) (18)
F2(x) = Y2(x) (19)
Where Y1(x) and Y2(x) are functions of the models corresponding to catalyst selectivity and reactor temperature respectively.
Therefore, in MOGWO both F1(x) andF2(x)havetobeminimizedinordertominimizereactortemperatureandmaximizecatalyst
selectivity.
5. Results and discussions
5.1. Performance of FNN-GWO model
The main goal of this research is to develop a closed model equation foran EOreactorthatisaccurate,simple,portable,andeasy
to understand.
The three-year dailyaverage dataset was cleaned, and 1220 data sets were qualified formodel construction after outlierswere
removed. A trial-and-error approach and a literaturestudy were used to determine the values of ANN parameters required for
modelling. The number of nodes in the hidden layer in the current study varies systematically from 5 to 25, and the FNN GWO
method is employed each time to discover the best weightsand bias that generatethelowestMSEbetweenactualandpredicted
output. These procedures give 21 [each model for each number of nodes] different ANN model with varying accuracy.
Shortlisting the models: The following criteria were used to pick a viable model from a pool of probable candidates or
representative model equations with varied degrees of complexity and accuracy:
(i) Simplicity: The model should have as little complexity as feasible. The number of nodesinthehiddenlayerwasusedto
determine model complexity.

(ii) Accuracy of prediction: The difference between expected and actual selectivity.
(iii) The model equation should capture the process's underlying physics. In other words, model equations should contain
physical understanding of the system under study, not just a predictive association. This is a crucial factor to
consider while creating realistic reactor models. Their plant operating knowledge, experienceandobservationsof
reactor behavior are summarized in Table 2.
TABLE 2 Rules to select best model from real Plant operation observations.
Sl. no Parameters changed keeping all other
parameters constant
What happen to
selectivity?
What happen to
temperature?
1 If oxygen inlet concentration increase Increase Decrease
2 If ethylene inlet concentration increase Increase Decrease
3 If CO2 inlet concentration increase Decrease Increase
4 If CG Pressure increase Increase Decrease
5 If CG flow increase Increase Decrease
6 If Work rate increase Decrease Increase
7 If Cum EOE/m3 of catalyst increase Decrease Increase
FIGURE 6: Actual vs. predicted plots of (a) catalyst selectivity with training data (b) catalyst selectivity with testing data (c)
reactor temperature with training data (d) reactor temperature with testing data

All 21 modelsacquired using ANN are subjected to the aforementioned inspection to determinewhetherthecreatedmodelisin
agreement with plant data. All of the produced models have been thoroughly tested. Assume aten-testdatasetwasconstructed
with all variables set to their 50th percentile value, with the exception of oxygen inlet concentration, which was altered in 10-
point intervals from its minimum to maximum value. The charts of oxygen inlet concentration vs. selectivity and oxygen inlet
concentration vs. temperature were created after that. Table 2 observation number 1 wasconfirmedusingthesecharts.Models
that do not match the findings in Table 2 are discarded because they do not represent the basic physics of the EO reactoranddo
not correlate with plant observations. They are essentially representations of a complicated data fitting model that makes no
sense.
Only one model forcatalyst selectivity and one model forreactor temperaturewerechosenfromtheshortlistedmodelsbecause
they are exceedingly precise,obey the Table 2 observations, andcapturetheinternalphysicsofthereactor,suchtwomodels are
considered the representative model equations for selectivity and temperature.
The corresponding Coefficient of Determination (R2) andAverage Percentage Error (APE) of the above models fortrainingand
test data have been mentioned in Table 3.
TABLE 3. Performance of ANN model
Training Testing
Model R2 APE R2 APE
Catalyst selectivity model 0.99 0.56 0.99 0.56
Reactor temperature model 0.99 0.24 0.99 0.23
With highR2 andlow APE forboth the catalyst selectivityand reactor temperaturemodels(Table3),itcanbeconcludedthatthe
predicted output values are comparable to the actual output values and that the models developed are reliable, fairly accurate,
and capture the inherent physics of the EO reactor. The model's generalizability and accurate learning on nonlinear input and
output relationships are further indicated by the strong R2 value on unseen test data and low APE.
Figure 6 depicts the models' prediction performance on training and testing data. The fact that the real and forecast curves
almost coincide implies that the model has strong prediction accuracy.
From Table 3 and Figure 6 it is concluded that developed model is highly accurate and reliable as it also performs well with
unseen test data.
Generation of explainable model equations: Though ANN generatesaclosedformofequationwhichhasveryhighpredictive
capability, the developed equation is large and complex and sometimes difficult to directly interpret. In present study, a
methodology is developed to enhance the interpretability of the developed equations. Figure 7 and Figure 8 summarize the
developed methodology. Those figures are generated by changing one variable at a time from its minimum to maximum value
(10 steps) while keepingall other 7 input variables at their 50-percentile value. Selectivity equations developedbyANNisused
to predict the selectivity value in each case of these simulated test data. After plotting was done,a trendline was drawnthrough
each data whose equationand R2 value is shown in figure. Based on visual inspection and R2 value trend line curvewasselected
(like straight line, or polynomial with degree 2 or 3 or more) so that generated trendline almost matches with the data. As seen
from the figure 10, the developed trend lines are very decisive and monotonically increasing and decreasing. As mentioned
earlier, they all match the plant actual observations and obey the Table 2. In short, developed models captures the nonlinear
relationship between selectivityandreactoroperatingparameters.Thesetrendlinescanbeusedbyplantoperatingengineersto
get the insights how a particular input parameter affects the catalyst selectivity. For example, from figure 7, it is quantitively
clear that increasing inlet oxygen and ethylene concentration, actually enhance the catalyst selectivity whereas increasing CO2
reduces the selectivity.
The selectivity increases linearly with oxygen concentration with positive slope of 0.6772, whereas the relation of selectivity
with ethylene and CO2 are nonlinear and represented by second order polynomial. Now these trend line equations are used to
develop the following explainable equation (Equation 20 and 21).

(20)
(21)
Where x1, x2, ……x8 are the actual value of the 8 input variables and x1, avg, x2, avg, ……x8, avg are the average (50 percentile) value of
input variables, respectively.
Each term in the equation 20 represents the change in selectivity if a particular parameter deviates from its average value. For
example, the term, represents the deviation ox inlet oxygen concentration from its average value and when it
multiplied by co-efficient 0.6772, represents the selectivity gain (or penalty) due to oxygen. In this way, all 8 parameters
contribution is calculated in equation 8 and it is added with 84.86% (average selectivity) to get the actual selectivity.
Main advantage of this equation (equation 20 and 21) over ANN model is that this equation is interpretable and easily
explainable to plant engineer. Equation is simple and contains terms or parametric co-efficient which throw lights relative
importance of each parameter on the overall selectivity if they deviate from this base value. Also, it indicates whetherthe effect
of each parameter is linear or non-linear.
Equation 20 is then used to predict selectivity of three-year actual data and predicted, and actual selectivity is compared.
Prediction error is 0.8% and R2 is 0.97. This low value of prediction error and high value of R2 signifies that developed equation
(equation 20) is highly accurate and reliable.
FIGURE 7: Influence of each variable on selectivity

FIGURE 8: Influence of each variable on reactor temperature
5.2. Optimization
5.2.1. Optimization through MOGWO
Once the reliable models were successfully developed, themodels were subjected to optimization.Purposeofoptimizationisto
find the optimum value of rector operating parameters to achieve maximum selectivity and minimum temperature
simultaneously. One of the critical tasks for optimization of any process is fixing of search space at which the optimal process
conditions are to be found out. Therefore, before running the optimization a lower bound and upper bound of the process
variables were fixed in consultation with plant engineers. As plant conditions are dynamic in nature, the total one operational
year was divided into three periods (SOR, MOR and EOR)and at each running period, optimizationswereperformed.Thelower
bounds (LB) and upper bounds (UB) considered in these three cases have been depicted in Table 4.
Table 4. Lower bounds and upper bounds for optimization
x1 x2 x3 x4 x5 x6 x7 x8
Case 1 : SOR
LB 3377 6.37 27.41 1.32 0.15 18.94 5.13 3455.62
UB 3387 9.00 35.12 2.50 0.24 19.95 9.22 4585.31
Case 2: MOR
LB 5377 6.37 27.41 1.52 0.15 18.94 6.13 3455.62
UB 5387 9.00 35.12 2.50 0.24 19.95 9.22 4585.31
Case 3: EOR
LB 8377 6.37 27.41 1.82 0.15 18.94 7.63 3455.62
UB 8387 9.00 35.12 2.50 0.24 19.95 9.22 4585.31

With the help of MOGWO code in MATLAB, a pareto diagram was developed forcase2(MOR)(Fig.8).Allthepointsinthe pareto
curve in Figure. 9 represents the pareto non-dominated optimal solution in catalyst selectivity and reactor temperature space
and all Pareto optimal solutions on the curve are considered equally good. As seen from the curve, if someone try to increase
selectivity, he must sacrifice reactor temperature. Reverse is also true, i.e., lower temperature can be achieved at the cost of
lower selectivity. Other than these solutions, there is no process condition which can increase selectivity without sacrificing
temperature and vice versa. Some of the pareto optimal solutions have been also tabulated in Table 5 for reference. From the
output of MOGWO the optimumvaluesofinputparameterscanbeobtainedwhichwillmaximizetheselectivityandminimizethe
temperature. Four solutions in table 5 are plotted as four small black circles on Figure 9 and represents trade-off between two
conflicting objectives. All the four solutions are equally good, and it is up to the plant operation engineer which criteria among
selectivity or temperaturethey will give priority. Based onthat, they can choose the optimuminputparametersfromthePareto
optimal solution. For example, if selectivity has more priority than reactor temperature for a particular plant, then operators
should choose solution number 4 in Table 5 which will help to achieve 86.1% selectivity by sacrificing the temperature(241.53
°C). On the other hand, if the temperaturehas more priority than selectivity then solution number 1 in Table4shouldbechosen
which will give a lower temperature by sacrificing selectivity. If temperature and selectivity have equal priority, then solution
number 2 of Table 5 should be chosen.
The main advantage of such a study is that it gives the operation engineer/ DCS panel operators a strategy to run the reactor in
optimum condition in real-time. In a running plant, since operators have no idea about Pareto optimal solution,operatorstryto
optimize the plant heuristically based on their experience and knowledge. One such real-life operating point is plotted and
named as ‘A’ in Figure. 9. As the operating point, A is not on the Pareto curve, itisnottheoptimumpointandimprovement scope
exists both in selectivityand temperature. From this plot inFigure.9,itcanbeconcludedthattheoperatorcanimprovetheplant
operatingconditions in two ways: selectivity can be improved keeping the temperature sameas before,orreactortemperature
can be decreased keeping thecatalyst selectivity as it is as shown by two arrows onFigure9.Theonlyactionoperatormustdois
to run the MOGWO with proper bound in real-time, and MOGWO will provide a set of optimum operating conditions that the
operator needs to set in the plant. This is the potential improvement area suggested by the Pareto optimal solution.
FIGURE 9. Pareto diagram of catalyst selectivity vs. reactor temperature
Table 5. Pareto optimal solution
x1 x2 x3 x4 x5 x6 x7 x8 Selectivity Temperature
7.75 18.25 0.31 22.64 782.15 161.09 2572.84 2.59 85.40 239.37
7.75 28.09 0.31 22.54 785.62 161.08 2572.84 2.52 85.71 240.01
7.74 28.18 0.31 22.17 785.93 161.09 2572.84 2.52 85.92 240.82
7.74 28.71 0.31 21.86 785.96 161.07 2572.84 2.53 86.10 241.53
A

6. Conclusion
From the existing plant operating data, this study uses Artificial Neural Networking to construct an accurate model of a
commercial ethylene oxide reactor. An ANN creates a closed model equation that is portable and may be used in a plant control
system. The true value of this research is that it has produced an explainable model equation that is very accurate and provides
insights into the process. Theproduced model equationsare based on the underlying physics of the process and areinlinewith
the observations and experiences of the plant engineer. After that, the developed model equations are used to construct multi-
objective pareto optimum solutions that optimize selectivity while lowering reactor temperature and thus ensure profit
maximization.
REFERENCES
[1] Lahiri, S.K., 2020. Profit Maximization Techniques for Operating Chemical Plants. John Wiley & Sons.
[2] Lakshminarayanan, S., Fujii, H., Grosman, B., Dassau, E., Lewin,D.R.,2000.Newproduct designvia analysisofhistorical
databases. Comput. Chem. Eng. 24, 671–676. https://doi.org/10.1016/S0098-1354(00)00406-3
[3] Grosman, B., Lewin, D.R., 2002. Automated nonlinear model predictive control using genetic programming. Comput.
Chem. Eng. 26, 631–640. https://doi.org/10.1016/S0098-1354(01)00780-3
[4] Searson, D., Willis, M., Montague, G., 2007. Co-evolutionof non-linearPLSmodel components.J.Chemom.21,592–603.
https://doi.org/10.1002/cem.1084
[5] Barati, R., Neyshabouri, S.A.A.S., Ahmadi, G., 2014. Development of empirical modelswithhighaccuracyforestimation
of drag coefficient of flow around a smooth sphere: An evolutionary approach. Powder Technol. 257, 11–19.
https://doi.org/10.1016/j.powtec.2014.02.045
[6] Floares, A.G. and Luludachi, I., 2014. Inferring transcription networks from data. In Springer Handbook of Bio-
/Neuroinformatics (pp. 311-326). Springer, Berlin, Heidelberg.
[7] Horne, B.G., 1993. Progress in supervised neural networks. Signal Process. Mag. IEEE 10, 8–39.
[8] Hagan, M.T., Menhaj, M.B., 1994. Training Feedforward Networks with the Marquardt Algorithm. IEEE Trans. Neural
Networks 5, 989–993. https://doi.org/10.1109/72.329697
[9] Adeli, H., Hung, S.L., 1994. An Adaptive Conjugate Gradient Learning Algorithm. Appl. Math. Comput. 62, 81–102.
[10] Zhang, N., 2009. An online gradient method with momentum for two-layer feedforward neural networks.
Appl. Math. Comput. 212, 488–498. https://doi.org/10.1016/j.amc.2009.02.038
[11] Zhang, J.R., Zhang, J., Lok, T.M., Lyu, M.R., 2007. A hybrid particle swarm optimization-back-propagation
algorithm for feedforward neural network training. Appl. Math. Comput. 185, 1026–1037.
https://doi.org/10.1016/j.amc.2006.07.025
[12] Gori, M., Tesi, A., 1992. On the problem of local minima in backpropagation. IEEE Trans. Pattern Anal. Mach.
Intell. https://doi.org/10.1109/34.107014
[13] Shaw, D., Kinsner, W., 1996. Chaotic simulated annealing in multilayer feedforward networks. Can. Conf.
Electr. Comput. Eng. 1, 265–269. https://doi.org/10.1109/ccece.1996.548088
[14] Koh, C.S., Hahn, S.Y., 1994. Detection of Magnetic Body using Artificial Neural Network with Modified
Simulated Annealing. IEEE Trans. Magn. 30, 3644–3647. https://doi.org/10.1109/20.312730
[15] Montana, D.J., Davis,L.,1989.TrainingFeedforwardNeural NetworksUsingGeneticAlgorithms.Proc.11thInt.
Jt. Conf. Artif. Intell. - Vol. 1 89, 762–767.
[16] Kiranyaz, S., Ince, T., Yildirim, A., Gabbouj, M., 2009. Evolutionary artificial neural networks by multi-
dimensional particle swarm optimization. Neural Networks 22, 1448–1462.
https://doi.org/10.1016/j.neunet.2009.05.013

[17] Settles, M., Rylander, B., 2002. Neural network learning using particle swarm optimizers. Adv. Inf. Sci. Soft
Comput. 224–226.
[18] Zhang, C., Li, Y., Shao, H., 2000. A new evolved artificial neural network anditsapplication.Proc.WorldCongr.
Intell. Control Autom. 2, 1065–1068. https://doi.org/10.1109/wcica.2000.863401
[19] van den Bergh, F., Engelbrecht, A.P., 2000. Cooperative Learning in Neural Networks using Particle Swarm
Optimizers. South African Comput. J. 26, 84–90.
[20] Zhang, C., Shao, H., Li, Y., 2000. Particle swarm optimization for evolving artificial neural network. Proc. IEEE
Int. Conf. Syst. Man Cybern. 4, 2487–2490. https://doi.org/10.1109/icsmc.2000.884366
[21] Mirjalili, S., Sadiq, A.S., 2011. Magnetic OptimizationAlgorithmfortrainingMultiLayer Perceptron.2011IEEE
3rd Int. Conf. Commun. Softw. Networks, ICCSN 2011 42–46. https://doi.org/10.1109/ICCSN.2011.6014845
[22] Si, T., Hazra, S., Jana, N.D., 2012.Artificial neural network trainingusingdifferential evolutionaryAlgorithmfor
classification. Adv. Intell. Soft Comput. 132 AISC, 769–778. https://doi.org/10.1007/978-3-642-27443-5_88
[23] Leung, Y., Gao, Y., Xu, Z. Ben, 1997. Degree of population diversity - A perspective on premature convergence
in genetic algorithms and its Markov chain analysis. IEEE Trans. Neural Networks 8, 1165–1176.
https://doi.org/10.1109/72.623217
[24] Hrstka, O., Kučerová, A., 2004. Improvements of real codedgeneticalgorithmsbasedondifferential operators
preventing premature convergence. Adv.Eng.Softw.35,237–246.https://doi.org/10.1016/S0965-9978(03)00113-3
[25] Liang, J.J., Qin, A.K., Suganthan, P.N., Baskar, S., 2006. Comprehensive learning particle swarm optimizer for
global optimization of multimodal functions. IEEE Trans. Evol. Comput. 10, 281–295.
https://doi.org/10.1109/TEVC.2005.857610
[26] Gao, W. feng, Liu, S. yang, Huang, L. ling, 2012. Particle swarm optimization with chaotic opposition-based
population initialization and stochastic search technique. Commun. Nonlinear Sci. Numer. Simul. 17, 4316–4327.
https://doi.org/10.1016/j.cnsns.2012.03.015
[27] Moslemipour, G., Lee, T.S., Rilling, D., 2012. A review of intelligent approaches for designing dynamic and
robust layouts in flexible manufacturing systems. Int. J. Adv. Manuf. Technol. 60, 11–27.
https://doi.org/10.1007/s00170-011-3614-x
[28] Mirjalili, S., Mirjalili, S.M., Lewis, A., 2014. Grey Wolf Optimizer. Adv. Eng. Softw. 69, 46–61.
https://doi.org/10.1016/j.advengsoft.2013.12.007
[29] Faris, H., Aljarah, I., Al-Betar, M.A., Mirjalili, S., 2018. Grey wolf optimizer: a review of recent variants and
applications. Neural Comput. Appl. 30, 413–435. https://doi.org/10.1007/s00521-017-3272-5
[30] Saremi, S., Mirjalili, S.Z., Mirjalili, S.M., 2015. Evolutionary population dynamics and grey wolf optimizer.
Neural Comput. Appl. 26, 1257–1263. https://doi.org/10.1007/s00521-014-1806-7
[31] Nadimi-Shahraki, M.H., Taghian, S., Mirjalili,S.,2021.Animprovedgrey wolfoptimizerforsolving engineering
problems. Expert Syst. Appl. 166, 113917. https://doi.org/10.1016/j.eswa.2020.113917
[32] Mirjalili, S., Saremi, S., Mirjalili, S.M., Coelho, L.D.S., 2016. Multi-objective grey wolf optimizer: A novel
algorithm for multi-criterion optimization. Expert Syst. Appl. 47, 106–119.
https://doi.org/10.1016/j.eswa.2015.10.039
[33] Pal, S., Chowdhury, S., Hens, A., & Lahiri, S. K. (2022). Artificial intelligence based modelling and multi-
objective optimization of vinyl chloride monomer (VCM) plant to strike a balance between profit, energy utilization
and environmental degradation. Journal of the Indian Chemical Society, 99(1), 100287.

Grey wolf optimizer trained ANN technique for development of explainable model of commercial ethylene oxide reactor and multi- objective optimization to maximize profit

More Related Content

Similar to Grey wolf optimizer trained ANN technique for development of explainable model of commercial ethylene oxide reactor and multi- objective optimization to maximize profit (20)

More from IRJET Journal (20)

Recently uploaded (20)

Grey wolf optimizer trained ANN technique for development of explainable model of commercial ethylene oxide reactor and multi- objective optimization to maximize profit