=Paper=
{{Paper
|id=None
|storemode=property
|title=An Optimisation Strategy for the Catalytic Transformation of Bioethanol into Olefins Using Computational Intelligence
|pdfUrl=https://ceur-ws.org/Vol-1422/115.pdf
|volume=Vol-1422
|dblpUrl=https://dblp.org/rec/conf/itat/SorrosalMBMA15
}}
==An Optimisation Strategy for the Catalytic Transformation of Bioethanol into Olefins Using Computational Intelligence==
https://ceur-ws.org/Vol-1422/115.pdf
J. Yaghob (Ed.): ITAT 2015 pp. 115–120
Charles University in Prague, Prague, 2015
An Optimisation Strategy for the Catalytic Transformation of Bioethanol into
Olefins Using Computational Intelligence
Gorka Sorrosal1 , Cristina Martin, Cruz E. Borges, Ana M. Macarulla, and Ainhoa Alonso-Vicario
Deusto Institute of Technology – DeustoTech Energy,
University of Deusto, Avda. Universidades 24, 48007 Bilbao, Spain,
gsorrosal@deusto.es,
http://energia.deusto.es
Abstract: This paper presents a strategy for the optimi- The optimal control laws of biorefinery production pro-
sation of the operational conditions of the catalytic trans- cesses are mainly unknown. Therefore, it is necessary to
formation of Bioethanol into Olefins (BTO) process. The test several operational conditions over the whole opera-
variables to optimise are the main operating variables of the tional range to study the influence of each manipulated
process (temperature, space-time and water content in the variable on the final production objectives.
feed), and the objective function is to maximise the total One of the key points for the implementation of the
production of olefins. The proposed strategy is based on BTO process is to perform an advanced control strategy
evolutionary algorithms guided by surrogate models used to by adjusting the operating variables to maintain product
simulate the process behaviour under different experimen- quality while extending the lifespan of the catalyst. Due
tal conditions. This paper compares the optimisation results to the influence of multiple variables simultaneously over
of the BTO process obtained using an existing mechanistic the reaction kinetics and the catalyst deactivation, it is nec-
model with those obtained with a surrogate model. The essary to develop advanced optimisation strategies of the
results suggest that the proposed methodology achieves operational conditions that guarantee specific production
similar results than those using mechanistic models but objectives without exceeding the operation limits to avoid
43 times faster. This is a preliminary study where only an irreversible deactivation of the catalyst.
constant set points have been tested; further research will Therefore, the search of the parameters and operational
include dynamic optimisation of the operational conditions conditions that allow to reach those production objectives
by testing expected dynamic trajectories for each operating give rise to optimisation problems in which the calcula-
variable. tion of an analytic solution can present some difficulties
with conventional search techniques [19]. While these tech-
1 Introduction niques require characteristics of the process or from the
optimisation problem, such as gradients, Hessians or linear-
Nowadays, we are becoming aware that crude oil is a finite ities, to calculate the next points; there are stochastic search
source of energy and raw material. Therefore, our society techniques as the Evolutionary Algorithms (EA) that solve
begins to impulse the sustainable development using alter- the optimisation problems only with stochastic rules.
native sources of energy and raw materials, such as coal In the field of chemical engineering, there are several
or biomass. In nature, being 170 billions tones of biomass applications where these algorithms have been employed
annually produced, only the 3-4 % is exploited. Thus, there for the design, optimisation and optimal control of chemi-
is a huge quantity of biomass available for its valorization cal reactors and plants [8, 1], such as in fermentation pro-
as raw material to obtain biofuels and other chemical prod- cesses with fed-batch reactors [19] or to optimise the opera-
ucts [6]. Consequently, the scientific developments in this tional conditions of industrial scale reactor [18] or chemical
field are very important in order to advance in a future plants [16].
post-petroleum society and to reduce our dependency from To perform a dynamic optimisation of a process or reac-
the petroleum and its derivatives. The development of new tor, one of the problems is that each cost function evalua-
tools to study the optimal operation of biomass transforma- tion may require from minutes to hours of calculation time,
tion processes for a future scaling up to industrial level is a and when using EA hundreds of evaluations are generally
new interesting research line. needed [14]. Therefore, as knowledge models use to be
An important biomass transformation process is the non linear (and hence difficult to be solved, even numer-
Bioethanol-To-Olefins (BTO) process. The use of biomass ically), surrogate models are commonly used to simulate
as raw material has a great interest as an alternative to the real process during its optimisation [13, 10]. Thanks
the petrochemistry for the production of light olefins like to their characteristics, Artificial Neural Networks (ANN)
ethylene and propylene. The use of the computational intel- are being increasingly used as a modelling technique for
ligence in this research field can improve the optimisation process simulation using evolutionary optimisation tech-
procedures and allow faster developments in the design and niques [2, 5].
optimal operation of the production processes. ANN attempt to mimic the structural principles of the
116 G. Sorrosal, C. Martin, C. E. Borges, A. M. Macarulla, A. Alonso-Vicario
biologic brains to learn the existing relations inside input- number of possible cycles of production-regeneration is
output datasets. They have been able to successfully model also limited. Depending on the operational conditions, both
any type of complex models and chemical reactors, such the production and the catalyst deactivation reversibility
as batch reactors [7, 12], laboratory or industrial scale re- will be affected.
actors [15] or even catalytic reactions as in the case of the Therefore it is necessary to explore experimentally those
BTO process [11]. operational conditions that fulfill the desired objectives,
The present work has the objective of performing an which would be very costly and complicated. The use
optimisation of the Bioethanol-To-Olefins (BTO) process of EA and surrogate models to explore those operational
in order to maximise the olefins total production while ex- conditions is much more practical and reasonable.
tending the catalyst lifespan. The BTO process, as other The following variables are the main operating variables
biomass transformation processes, uses a specific catalyst that govern the behaviour of the BTO process:
to stimulate the formation of a specific product at the reac-
tor output. This catalyst is deactivated with time depending • Operating variables:
on the operational conditions. Therefore, in the design
and optimisation of new catalytic transformation processes, T : reaction temperature (K).
such as the BTO process, there are two aspects to take into Xw : mass fraction of water based on the equivalent
account in order to maximise the production of the desired mass of ethylene in the reactor feed gwater g−1 .
product: the composition of the catalyst and the operational −1
conditions. In the first case, different approaches based W FEO : space-time gcatalyst h−1 (gethanol ) .
on soft computing techniques have been proposed to op-
• Activity level:
timise the catalyst composition [17]. For the operational
conditions, we can find some studies for different processes a: catalyst activity. The activity has been consid-
but without catalyst deactivation [16] or only in a discrete ered as a disturbance that quantifies the rate of
way [9]. In this work, we proposed a strategy to study the catalyst deactivation by coke.
optimal dynamic operational conditions (with a specific
catalyst composition) to achieve the optimal production In previous works, experimental runs of this process
results, taking into account the catalyst deactivation to de- were carried out in an automated device equipped with an
sign an optimal operation policy that not only maximise isothermal laboratory scale fixed bed reactor connected on-
the production objectives but it is also able to counteract line to a gas chromatograph and a micro-GC for the analysis
the catalyst deactivation. of the reaction product (see Figure 1). Details about the
This paper presents the preliminary results for the opti- reaction equipment, catalyst preparation and experimental
misation of constant operation set points. The optimisation methodology can be found in previous works [3, 4].
objectives are the operating conditions that govern the BTO
process. The optimisation process has been implemented TIC
using computational intelligence algorithms. An EA ex- PC
plores the possible optimal solutions guided by a surrogate
model of the process based on ANN that is integrated in its
Liq. feed TIC
FT1 FIC1
evaluation function. N2/He
The paper is organized as follows: Section 2 presents FT2 FIC2
the BTO process. Next, Section 3 describes the proposed Air
methodology for the optimisation of the BTO process. Sec- FT3 FIC3
tion 4 presents the obtained results. The conclusions and Aux. gases
future works are finally summarized in Section 5. FT4 FIC4
H2/He
2 BTO Process H2/He
LIC
exhaust
The BTO process consists in the catalytic transformation
of bioethanol into olefins over an acid catalyst. This is a
key process in the concept of sustainable refinery, incor- Figure 1: Diagram of the reaction equipment used to obtain
porating biomass or derivatives as an alternative feedstock the experimental data.
to petroleum. This process uses a very selective catalyst
to maximise the olefins conversion rate (XO ). However,
this catalyst is deactivated due to the accumulation of coke.
Therefore, when the catalyst reaches a minimum activity 3 Methodology
value, it is necessary to stop the production step and carry
out the catalyst regeneration phase. In addition, the regen- This section presents the proposed dynamic optimisation
eration does not achieve the total catalyst recovery, so the strategy using Soft Computing techniques. An EA guided
An Optimisation Strategy for the Catalytic Transformation of Bioethanol into Olefins Using Computational Intelligence 117
by an ANN based surrogate model of the process is pro- A first ANN model, trained only with experimental data,
posed. The surrogate models are used in the evaluation op- was able to describe correctly the process behaviour for the
erator to simulate the process behaviour under each opera- experimentally tested operating conditions (see Section 4).
tional condition proposed by the EA. The results will be val- However, for optimisation purposes, it is necessary to test
idated with an existing mechanistic model (MECH) [3, 4], several operating conditions where limited experimental
since the experimental validation is not technical or eco- data are available. In particular, the dataset does not contain
nomically viable. any experiment describing the dynamic behaviour of XW
−1
The EA and the BTO process models (both the MECH and W FEO . To provide the ANN with the required infor-
and the surrogate) have been implemented using the pro- mation about the process behaviour in the whole operating
gramming package MATLABTM (version 8.0, 2012b, Math- range, some experiments were simulated with the MECH
works Company). Simulation and optimisation have been model and introduced in the training dataset.
carried out in an PC with an Intel R CoreTM i5-2467M
CPU at 1.6 GHz and 4.0 GB of physical memory (RAM).
3.2 Optimisation Problem
3.1 Surrogate Model The aim of this optimisation problem is to obtain the oper-
ational conditions that achieve the best production results
Chemical knowledge models are generally computationally per amount of catalyst needed. Due to the catalyst deacti-
very demanding to be used in evolutionary approaches [14]. vation, is important to bear in mind the deactivation rate,
Thus, in this work, an ANN based surrogate model is pro- keeping in the whole simulation time the catalyst activity
posed to dynamically simulate the process behaviour. and the olefins production rate over the established mini-
A nonlinear autoregressive with exogeneous inputs mum value of 0.10. Therefore, the Equation (2) defines the
(NARX) neural network topology has been selected to objective function in order to maximise the total production
model the BTO process. The developed ANN model esti- of olefins.
mates the olefins conversion rate at the reactor output (XO ) Z τ
using as inputs the previously mentioned process operat- −1
−1 XO T, Xw ,W FEO dt
ing variables (T , XW ,W FEO ), the catalyst activity level (a) max 0
. (2)
−1
and the previously estimated output values in a recursive −1
T,Xw ,W FEO W FEO
loop (X̂O ).
An iterative methodology modifying the number of lay- Please note that there are two stopping criteria (τ constant
ers and neurons in each layer has been carried out in order in the upper bound of the integral):
to select the neural model structure that better fits the pro-
cess using the Leave-One-Out Cross-Validation (LOOCV) • The catalyst involves a major process cost. Therefore,
technique. This technique consists of setting aside a set of to maximise the production per amount of catalyst
experiments (representing unique operational conditions) we have set a lower bound on the activity in order to
from the model training phase and only using them for reduce the total production costs. This bound has been
the validation phase. The process is repeated until every set to 0.10 (a < 0.10).
single set of experiments is used in the validation stage. • In order to maintain the production, it is needed that
These type of techniques test the generalization capability the olefins conversion rate does not decrease below
of a model structure. the 0.10 (XO < 0.10).
Consequently, the available data are divided into training,
validation and test datasets. The training and validation If any of the above criteria are passed over, we consider that
datasets are used in the model structure selection. The the production has reached its maximum span and should
first ones are used to train several models with different be stopped to proceed with a catalyst regeneration phase.
structures. Aspects such as the number of hidden layers, the The operating variables to optimise are bounded based
neurons in each layer or the connections between neurons on the physical-chemical properties of the process. In
are iteratively modified. The performance of each model is fact, for the temperature (T ), 573K is the inferior bound
tested and compared following the LOOCV procedure. where the complete dehydration of the ethanol happens
Once the neural model structure is selected, the final and 673K is the upper bound to avoid the irreversible de-
model is trained with all the available data except from activation of the catalyst. The variable XW will range
the test dataset which is used to validate the fitted model. between [0.0821, 4.8889] and W FEO −1
will range between
The Levenberg-Marquardt Algorithm has been used for the [0.068, 1.525] respectively. Please note that those inter-
training and the main comparison criterion has been the vals have been chosen as are the ones used to adjust the
Root Mean Squared Error (RMSE) of the model (Equa- mechanistic model [3, 4].
tion (1)). As previously stated, in order to solve the optimisation
s problem an EA has been used. In particular we have im-
1 n plemented a Genetic Algorithm (GA). Table 1 summarizes
RMSE := ∑ (XO − X̂O )2 .
n i=1
(1)
the principal parameters of the EA used. The surrogate
118 G. Sorrosal, C. Martin, C. E. Borges, A. M. Macarulla, A. Alonso-Vicario
Parameter Value 1
Xw = 0.888 gwaterg−1 MECH
T = 673 K ANN
Individuals Vector of real numbers 0.8 W/FEO = 0.25 gcatalysth(gethanol)−1 a
Population 50 individuals Production (ANN) = 61.5311 go g−1
catalyst
Xo (gog−1)
0.6
Generations 200 generations Production (MECH) = 60.3583 go g−1
catalyst
Fitness Operator Olefins Production 0.4
Genesis Operator Random (uniform) initialization
Selection Operator 4-Tournament 0.2
Crossover Operator Arithmetic crossover
Mutation Operator Adaptive mutation 0
0 10 20 30 40 50 60 70
Elitism True time on stream (h)
Table 1: Main parameters of the GA used to optimise the Figure 3: Comparison of the estimation generated by the
operational conditions. MECH and ANN models for a test experiment.
model previously developed will be used now to simulate MECH (go g−1 ANN (go g−1
catalyst ) catalyst )
the temporal behaviour of the process under the operating
µProduction σProduction µProduction σProduction
constant set points provided by the GA.
70.2204 0.6385 63.7376 1.6946
4 Results Table 3: Comparison of the mean production results for the
“optimal” operational conditions when repeating several
In this section the results obtained for the modelling and times the evolutionary optimisation using the mechanistic
optimisation of the BTO process are shown. Following (MECH) and surrogate (ANN) models.
the modelling procedure mentioned above, an ANN based
surrogate model, with the topology showed in the Figure 2,
Figure 3 represents the estimates of the process be-
has been implemented.
haviour calculated with both models for a test operational
conditions that were excluded for the training procedure.
The ANN model has been validated using the testing dataset
(see Section 3), obtaining a root mean squared error of
0.0323 go g−1 for the whole test experiments. These results
show the capacity of the ANN to properly assimilate and
reproduce the BTO process dynamics in the same way as
the mechanistic model.
Once the surrogate model has been validated, the evolu-
tionary optimisation has been carried out. Table 3 shows
the mean production results obtained with the “optimal”
operational conditions generated by the EA. Being the opti-
misation procedure stochastic, it has been launched several
times to guarantee its convergence to a local optimum. No-
Figure 2: Feed-Forward ANN topology for the BTO pro- tice the small deviations repeating all the procedure 50 and
cess model. 100 times for the mechanistic and surrogate approaches
respectively. Although the standard deviation using the
Table 2 shows the mean estimation errors of the mecha- surrogate model doubles the deviation that results from
nistic and surrogate models for the experimental data. The using the mechanistic model, in both cases they are still
discrepancy between both models is rather low, but the very small. So we can conclude that the proposed approach
surrogate model simulates the process behaviour 43 times is converging to a local optimum.
faster. Finally, Figure 4 shows the behaviour of the process
under the optimum solution provided by the evolutionary
Model RMSE (go g−1 ) optimisation using the surrogate model. The maximum pro-
duction has been 68.05 go g−1
catalyst . Please note that the max-
MECH 0.0323 imum production using the surrogate model only differs
ANN 0.0387 a 5.67 % over the optimisation carried out using the mech-
anistic model, but with much less computation cost. This
Table 2: Root mean square error of the mechanistic maximum is reached by operating the reactor at 645K with
(MECH) and surrogate (ANN) model to experimental data. a high content of water in the feed (XW = 4.875gwater g−1 )
An Optimisation Strategy for the Catalytic Transformation of Bioethanol into Olefins Using Computational Intelligence 119
1
Xw = 4.8753 gwaterg−1 [2] Farshad, F., Iravaninia, M., Kasiri, N., Mohammadi, T.,
Production
T = 645.6708 K Activity Ivakpour. J.: Separation of toluene/n-heptane mixtures
0.8 W/FE0 = 1.2933 gcatalysth(gethanol)−1
lim (0.10) experimental, modeling and optimization. Chem. Eng. J.
Production = 68.0554 gOg−1
catalyst 173(1) (2011) 11–18
XO (gOg−1)
0.6
[3] Gayubo, A. G., Alonso, A., Valle, B., Aguayo, A. T., Bil-
0.4 bao., J.: Kinetic model for the transformation of bioethanol
into olefins over a HZSM-5 zeolite treated with alkali. Ind.
0.2 Eng. Chem. Res. 49(21) (2010) 10836–10844
[4] Gayubo, A. G., Alonso, A., Valle, B., Aguayo, A. T., Bil-
0
0 100 200 300 400 500 bao., J.: Deactivation Kinetics of a HZSM-5 zeolite catalyst
time on stream (h) treated with alkali for the transformation of bio-ethanol into
hydrocarbons. American Inst. Chem. Eng. 58(2) (2012) 526–
Figure 4: BTO process performance under the best solution 537
provided by the evolutionary optimisation guided by the [5] Gueguim Kana, E. B., Oloke, J. K., Lateef, A., Adesiyan,
surrogate model. M. O.: Modeling and optimization of biogas production
on saw dust and other co-substrates using artificial neural
network and genetic algorithm. Renew. Energ. 46 (2012)
and a space-time of 1.293go h(gethanol )−1 . This water quan- 276–281
tity attenuates the catalyst deactivation, which allows the [6] Huber, G. W., Corma, A.: Synergies between bio- and oil
extension of the production phase. refineries for the production of fuels from biomass. Angew.
Chem. Int. Edit. 46(38) (2007) 7184–7201
[7] Kashani, M. N., Shahhosseini, S.: A methodology for mod-
5 Conclusions
eling batch reactors using generalized dynamic neural net-
works. Chem. Eng. J. 159(1–3) (2010) 195–202
An optimisation strategy for the catalytic transformation
[8] Kordabadi, H., Jahanmiri, A.: Optimization of methanol
of Bioethanol-To-Olefins (BTO) process based on com-
synthesis reactor using genetic algorithms. Chem. Eng. J.
putational intelligence has been presented. The proposed
108(3) (2005) 249–255
evolutionary optimisation guided by an Artificial Neural
[9] Kordabadi, H., Jahanmiri, A.: A pseudo-dynamic optimiza-
Networks based surrogate models has been able to opti-
tion of a dual-stage methanol synthesis reactor in the face
mise the process obtaining similar results than those ob- of catalyst deactivation. Chem. Eng. Process. 46(12) (2007)
tained when using a mechanistic model but 43 times faster. 1299–1309
A clear optimum has been defined. The optimum opera- [10] Laguna, M., Martí, R.: Neural network prediction in a sys-
tional conditions have shown to be, temperature: 645K; tem for optimizing simulations. IIE Trans. 34 (2002) 273–
water quantity in the feed: 4.875gwater g−1 ; space-time: 282
1.293go h(gethanol )−1 . These operational conditions allow [11] Molga, E. J.: Neural network approach to support modelling
to extend the catalyst lifespan maximising the production of chemical reactors: problems, resolutions, criteria of ap-
phase and hence the total production. plication. Chem. Eng. Process. 42(8–9) (2003) 675–695
The presented results are a preliminary study on the op- [12] Mujtaba, I. M., Aziz, N., Hussain, M. A.: Neural network
timisation of the BTO process looking only for constant based modelling and control in batch reactor. Chem. Eng.
set points that maximise the production objectives. This Res. Des. 84(8) (2006) 635–644
has been the first step for a more complex dynamic op- [13] Nascimento, C. A. O., Giudici, R., Guardani, R.: Neural net-
timisation of the process, in which the optimal dynamic work based approach for optimization of industrial chemical
trajectory for each operating variable will be described. In processes. Comput. Chem. Eng. 24 (2000) 2303–2314
this line, several trajectory types will be defined for each [14] Ong, Y. S., Nair, P. B., Keane, A. J., Wong, K. W.: Surrogate-
of the operating conditions to study the process dynamic assisted evolutionary optimization frameworks for high-
behaviour. fidelity engineering design problems. In: Knowledge In-
corporation in Evolutionary Computation, (2004), 307–331,
Springer Berlin Heidelberg
Acknowledgement [15] Rahimpour, M. R., Shayanmehr, M., Nazari, M.: Modeling
and simulation of an industrial ethylene oxide (EO) reactor
The research team would like to thank the financial sup- using artificial neural networks (ANN). Ind. Eng. Chem.
port of the Basque Government by the research project Res. 50(10) (2011) 6044–6052
BIOTRANS (PI2013-40). [16] Rajesh, J. K., Gupta, S. K., Rangaiah, G. P., Ray, A. K.: Mul-
tiobjective optimization of industrial hydrogen plants. Chem.
Eng. Sci. 56(3) (2001) 999–1010
References [17] Rodemerck, U., Baerns, M., Holena, M., Wolf., D.: Appli-
cation of a genetic algorithm and a neural network for the
[1] Angira, R., Babu, B.V.: Optimization of process synthe-
discovery and optimization of new solid catalytic materials.
sis and design problems: A modified differential evolution
Appl. Surf. Sci. 223(1–3) (2004) 168–174
approach. Chem. Eng. Sci. 61(14) (2006) 4707–4721
120 G. Sorrosal, C. Martin, C. E. Borges, A. M. Macarulla, A. Alonso-Vicario
[18] Yee, A. K. Y., Ray, A. K., Rangaiah, G. P.: Multiobjective
optimization of an industrial styrene reactor. Comput. Chem.
Eng. (2003) 111–130
[19] Yüzgeç, U., T urker, M., Hocalar, A.: On-line evolutionary
optimization of an industrial fed-batch yeast fermentation
process. ISA Trans. 48(1) (2009) 79–92