Numerical Simulation and Novel Construction Methods in Oil and Gas Engineering - Research Article

Advances in Mechanical Engineering

2019, Vol. 11(3) 1–12
Ó The Author(s) 2019
Optimization of vane demister based DOI: 10.1177/1687814019835105
journals.sagepub.com/home/ade
on neural network and genetic
algorithm

San He1 , Hang Liu1 , Yongli Zou1 and Qinqin Zhang2

Abstract
A vane demister is widely used for separating tiny droplets from gas streams in the petroleum industry, chemical engi-
neering, and other industries. To obtain optimal structure and operation parameters, a method based on orthogonal
experiment design is often adopted. However, in most cases, results from an orthogonal experiment design are subopti-
mal solutions when there are fewer experiments to optimize the vane demister performance. In this study, to obtain the
maximum separation efficiency and minimum pressure drop, Fluent software was used to simulate the two-phase flow of
gas and liquid in vane demister with different structural parameters and operation parameters, generating 473 solutions
as the sample database. Based on this database, a back propagation neural network was used to establish the prediction
model for the separation efficiency and pressure drop, and a genetic algorithm was used for multi-target optimization of
this model. The optimization results were compared to Fluent simulation results and the orthogonal experiment design
results. The results show that a genetic algorithm generates better results. The optimal separation efficiency of both
methods is 100%. However, the optimal pressure drop of the genetic algorithm is 25.77% lower than that of the ortho-
gonal experiment design.

Keywords
Vane demister, optimization, orthogonal experiment design, back propagation neural network, genetic algorithm

Date received: 26 August 2018; accepted: 4 February 2019

Handling Editor: Hongfang Lu

Introduction results in the separation of gas and liquid.1 A vane

demister can separate droplets with diameters larger
During natural gas processing, liquid impurities carried
by the natural gas can cause corrosion of pipelines and
equipment and therefore affect production. Gravity 1
State Key Laboratory of Oil and Gas Reservoir Geology and
separators are used to remove large droplets from the Exploitation, Southwest Petroleum University, Chengdu, P.R. China
2
gas stream in the production, whereas the removal of Sichuan College of Architectural Technology, Deyang, P.R. China
fine droplets is primarily performed by a demister. The
Corresponding authors:
vane demister is a commonly used demister because of San He, State Key Laboratory of Oil and Gas Reservoir Geology and
its good maintainability. It removes entrained fine Exploitation, Southwest Petroleum University, Chengdu 610500, P.R.
liquid droplets from a gas flow by inertial impinge- China.
ment. When the gas and entrained fine liquid droplets Email: [email protected]
flow through narrow channels in a vane demister, the
Hang Liu, State Key Laboratory of Oil and Gas Reservoir Geology and
gas can pass through freely, whereas the small droplets Exploitation, Southwest Petroleum University, Chengdu 610500, P.R.
cannot make the necessary sharp turns due to their China.
inertia. Small droplets strike the channel walls, which Email: [email protected]

Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License
(http://www.creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/
open-access-at-sage).
2 Advances in Mechanical Engineering

than 8 mm. It has the advantages of high separation pressure drop of the demister. The simulation results
efficiency, low pressure drop, a higher gas velocity, a show that more number of bends or smaller bend wave-
simple structure, and easy maintenance. It has been lengths result in a higher separation efficiency and a
widely used in the separation processes of washing higher pressure drop.
towers, cooling towers, evaporators, columns, and Currently, the CFD simulation is the main method
separators. for studying the vane demister. It has the advantages of
Presently, studies of the vane demister are focused shorter research time and lower cost. At the same time,
on the effect of different factors on the performance of the accuracy and reliability of the simulation can satisfy
the demister, selection of a turbulence model in a com- the engineering design requirement. Because the selec-
putational fluid dynamic (CFD) model, and optimiza- tion of the turbulence model affects the simulation
tion of the performance of the demister. results, CFD studies are focused on this selection. YI
The treatment capacity, separation efficiency, and Wang and PW James6 compared the CFD simulation
pressure drop are all parameters that can characterize results and experimental data of a wave plate demister
the performance of the demister. The treatment capac- and found that the predicted value of a low Reynolds
ity of the vane demister is determined by the critical number k-e turbulence model is more accurate than
speed at which the secondary droplets are generated.2 that of a standard k-e turbulence model. G Venkatesan
When the gas velocity exceeds a certain critical value, a et al.7 performed a numerical simulation on a curved
secondary carrying phenomenon occurs, that is, the vane demister using 15 different turbulence models and
separated droplets will escape from the demister outlet found that Spalart–Allmaras model and few variations
again with the airflow, eventually resulting in a decrease of k-e model showed better prediction results than other
in the separation efficiency. The mechanism of second- two equation models.
ary droplet generation is complicated, and the capacity Although CFD simulations are the main method for
calculation is also very complicated. However, the effi- studying the vane demister, an optimal combination of
ciency and pressure drop are easier to obtain. The parameters can hardly be obtained based only on the
separation efficiency indicates the separation ability of CFD simulation. This is because a CFD simulation
the vane demister. A higher separation efficiency means requires a significant amount of calculation and contin-
a lower liquid content in the gas after separation by the uous adjustment of the affecting parameters to obtain
demister, which better satisfies production require- optimal parameters. At the same time, design difficul-
ments. The pressure drop between the inlet and outlet ties can increase if the separation efficiency and pres-
indicates the energy consumption of the demister. A sure drop need to be considered and balanced. Hence,
higher pressure drop means that a greater energy supply the CFD simulation is mostly combined with an opti-
is needed after using the demister, which increases the mization method to obtain the optimal parameter com-
production cost. Therefore, the efficiency and pressure bination of the demister. J Zhao et al.8 conducted
drop are crucial parameters in characterizing the demis- numerical simulations on vane demisters with different
ter performance,3 and a demister with a high separation geometric shapes and operation conditions. The predic-
efficiency and a low pressure drop is preferred. tion model of separation efficiency was obtained based
There are many factors that affect the separation on the response surface method. This prediction model
efficiency and pressure drop of a vane demister. can predict the effect of different geometric shapes and
Related studies have been performed by many research- operation conditions on separation efficiency, and it
ers. SJ Xu et al.4 performed a numerical simulation on can be used to optimize the demister parameters based
a demister with streamlined wave plate and found that on the different separation requirements. E Narimani
the separation efficiency increases with the liquid dro- and S Shahhoseini9 studied the effect of gas velocity,
plet diameter and inlet velocity and decreases with the vane spacing, and vane angle on the separation effi-
increase in vane spacing. The pressure drop at the inlet ciency of a vane demister by CFD simulation. The pre-
and outlet is not affected by the diameter of the droplet. diction model of separation efficiency was obtained
However, it increases with the inlet gas velocity. With using a response surface method and the optimal para-
the increase in the vane spacing, the pressure drop first meters were determined. However, this prediction
decreases and then increases. X Wang5 studied the model was not used in the optimization calculation.
demister using numerical simulation and experiments. The result was a set of parameters in the designed
It was found that the flow field and pressure drop of experiments, and the prediction accuracy of the model
the demister were primarily affected by the windward on non-sample data was not discussed. G Venkatesan
surface structure parameters of the demister vane. et al.10,11 arranged 25 orthogonal experiments and per-
MHH Estakhrsar and R Rafee3 performed a numerical formed numerical calculations using Fluent software to
simulation on a wave plate demister with a drainage finally determine the optimal parameters. Based on the
channel and studied the effect of the number of bends results from the orthogonal experiment designs, var-
and wavelength on the separation efficiency and iance analyses were performed to rank the importance
He et al. 3

level of structure parameters on the friction factor and

droplet separation efficiency. The orthogonal experi-
ment design is an experimental design method with rel-
atively fewer experiments to determine the optimal
combinations of different factors. However, the real
optimal condition may not be from the designed experi-
ments. If the experiment design is not ideal and fine,
the optimal results can be very different from theoreti-
cal results.
Based on the above issues, a novel idea is applied in
this article for optimization of the vane demister. First,
many CFD simulations were performed and the results
were used as sample data. The prediction model of
separation efficiency and pressure drop was established
based on back propagation (BP) neural networks. The
model was optimized using a genetic algorithm to
obtain the optimal combination of demister para-
meters. CFD simulations were used to verify the mod-
el’s ability to predict non-sample data. Finally, the Figure 1. A vane demister with drainage channels.
results were compared to the traditional orthogonal (1) Vane spacing; (2) Vane height; (3) Vane angle; (4) Length of drainage
design results. channel.

Research methods The discrete phase model (DPM) was used to simulate
CFD simulation the demister internal flow field, and the Euler–
Lagrangian algorithm was used in the calculation.
With the rapid development of computer and CFD Furthermore, the effect of gravity and the coupling of
techniques, numerical simulation has become the main gas and liquid phases were also considered. The contin-
method for simulating the internal flow field of demis- uous phase is methane, and the discrete phase is a
ters. Compared to experimental methods, CFD simula- water drop with a particle diameter of 15 mm. The
tion has the advantages of shorter research time and quality of gridding directly affects the accuracy of the
lower cost. In demister studies, the CFD simulation results. Therefore, a top-down cutting method was used
results are in good agreement with the experimental to generate a high-quality quadrilateral structure grid,
results.9,12,13 In this study, CFD calculation replaced as shown in Figure 2. The quantity of the grid was con-
experiments, and Fluent software was used for numeri- trolled within 20,000 to 40,000.
cal simulation of a demister. In this study, CFD calcu-
lation replaced experiments and Fluent software was
used to simulate the demister to calculate separation
efficiency and pressure drop for different vane spacings,
vane angles, vane heights, drainage channel lengths,
gas velocities, and temperatures. The separation effi-
ciency is calculated by equation (1), and the pressure
drop is directly read from the Fluent software
mn
h= 3 100 ð1Þ
m
where m is the number of droplets at the inlet of the
demister, n is the number of droplets at the outlet of
the demister, and ó is separation efficiency (%).
The research objective of this article is shown in
Figure 1. Drainage channels are added where the flow
direction changes in the demister to obtain a better
separation effect. Due to the similarity of the flow con-
ditions between vanes, only one channel between the
two adjacent vane was selected as the simulation object. Figure 2. Grid cutting diagram.
4 Advances in Mechanical Engineering

Turbulence model. The internal flow field of a demister is

primarily based on swirl flow. The renormalization
group (RNG) k-e model considers the swirl effect and
is suitable for calculation of the strong swirl flow
model. Furthermore, RNG theory presents the analyti-
cal equation for the viscosity of low Reynolds number
flow. The analytical equation has high reliability and
accuracy in calculation of the near wall area flow.14
According to the research results of Wang and James,6
the low Reynolds number k-e turbulence model shows
ideal results in simulation of the demister internal flow
field. Therefore, the RNG k-e turbulence model is
selected in simulations.

Simplified conditions. According to the characteristics of

Figure 3. Independence analysis of grid.
the vane demister structure shown in Figure 1, the cal-
culation conditions were simplified as follows:
quantities, and scaled residuals for solution con-
1. The flow field is two-dimensional. vergence are set to be 10–4.
2. The gas is non-compressible.
3. The gas flow is steady.
4. There is no interaction between droplets, no colli- Grid independence analysis. To analyze the effect of grid
sion, no aggregation, and no secondary droplet. number on the simulation results, six different grids
5. Once the droplet touches the wall, it is consid- were generated to simulate the internal gas flow of the
ered removed and no bouncing is considered. vane demister with vane spacing 25 mm, vane angle
100°, vane height 180 mm, and drainage channel length
14 mm at conditions of gas velocity 4 m/s and tempera-
Boundary conditions. ture 100°C. The results in Figure 3 show that the
1. The inlet is the inlet of velocity and the outlet is separation efficiency and pressure drop are almost con-
the outlet of pressure. The wall and the drainage stant when the number of grids is more than 20,000.
channel were set as the boundary of the wall. Therefore, the number of grids for all models in this
The surface roughness is zero with no slip and article is controlled at more than 20,000.
thermal insulation.
2. The velocity inlet and pressure outlet were set as
escape; the wall and the drainage channel were Orthogonal experiment design
set at trap. In this article, six parameters of the vane spacing D,
3. Droplets have the same velocity as air at the vane angle a, vane height H drainage channel length L,
inlet. gas velocity v, and temperature T are selected as
4. The SIMPLE calculation method is used for research factors, and five levels were selected for each
coupling the pressure and velocity, the second- factor for the orthogonal experiment design. The
order upwind differencing scheme is used for selected factors and levels are listed in Table 1.
convective terms and for the turbulence According to the structural theory of the orthogonal

Table 1. Affecting factors and levels for the vane demister.

Series no. Parameter/factor Symbol Level of the parameters

1 2 3 4 5

1 Vane spacing (mm) D 20 25 30 35 40

2 Vane angle (°) a 90 100 110 120 130
3 Vane height (mm) H 160 170 180 190 200
4 Drainage channel length (mm) L 8 10 12 14 16
5 Gas velocity (m/s) v 2 3 4 5 6
6 Temperature (°C) T 90 100 110 120 130
He et al. 5

Table 2. Orthogonal array for the given parameters and levels.

Trail no. Factors and their levels

D a H L v T

E1 1 1 1 1 1 1
E2 1 2 2 2 2 2
E3 1 3 3 3 3 3
E4 1 4 4 4 4 4
E5 1 5 5 5 5 5
E6 2 1 2 3 4 5
E7 2 2 3 4 5 1
E8 2 3 4 5 1 2
E9 2 4 5 1 2 3
E10 2 5 1 2 3 4
E11 3 1 3 5 2 4 Figure 4. The structural model of a BP neural network.
E12 3 2 4 1 3 5
E13 3 3 5 2 4 1
E14 3 4 1 3 5 2
E15 3 5 2 4 1 3 another factor are selected as variables, and the remain-
E16 4 1 4 2 5 3 ing four factors are used as fixed values. And the velo-
E17 4 2 5 3 1 4 city varies from 1 to 7 m/s, and the range of the other
E18 4 3 1 4 2 5
E19 4 4 2 5 3 1 variable is not fixed. In the next batch, the velocity is
E20 4 5 3 1 4 2 still used as a variable, and another variable will be ran-
E21 5 1 5 4 3 2 domly selected from the remaining five factors. Then,
E22 5 2 1 5 4 3 the remaining four factors still take a fixed value, but
E23 5 3 2 1 5 4
the values of the same fixed factor are different in dif-
E24 5 4 3 2 1 5
E25 5 5 4 3 2 1 ferent batches.
The learning samples are randomly divided into
three sets, namely, the training set, the verification set,
and the test set. Training ratio is 0.7, validation ratio is
table, an orthogonal array of 25 cases (L25) for six 0.15, and test ratio is 0.15.
parameters with five levels was selected, as shown in
Table 2.
Parameter settings. It has been theoretically proven17
that for a three-layer BP neural network with an input
BP neural network modeling layer, hidden layer, and output layer, if the number of
An artificial neural network is a computing model that neurons in the hidden layer is large enough, it can infi-
simulates the human brain neural network during infor- nitely approximate any continuous function on a non-
mation processing. It has a strong self-adaptive and unbounded region. Hence, a three-layer BP neural net-
self-learning ability as well as a powerful non-linear work was used in this study. Six factors, D, a, H, L, v,
reflection ability that can accurately predict the target and T, were used as model input, whereas h and DP
value. Presently, the most often used method is the BP were used as output. The neuron number of the hidden
neural network. Also known as the error BP neural net- layer was determined as 13 using equation (2).18 Thus,
work, it is a multi-layered forward neural network that a 6-13-2 BP neural network structure was established,
has a strong non-linear prediction ability, generaliza- as shown in Figure 4
tion ability, and fault tolerance. Therefore, it has been
widely used for non-linear applications and approxima- Nhid = 2Nin + 1 ð2Þ
tions in engineering applications.15,16 In this study,
where Nhid is the number of hidden neurons and Nin is
neural network modeling was established using the
the number of input neurons.
neural network toolbox in MATLAB.
Because the units of the six input data are different,
the value span is large. If the input data are directly
Sample data. In this article, 448 sets of CFD simulation used as the training sample, it will not only increase the
results and 25 sets of orthogonal test data were used as training convergence time of the neural network but
learning samples for BP neural networks. The 448 sets also affect the accuracy of the network prediction
of data were derived from simulation results for several results.19,20 Therefore, the input data were normalized
batches. In each batch simulation, gas velocity and by the mapminmax function and the input values are
6 Advances in Mechanical Engineering

scaled to the interval of 0–1. The mapminmax function solution. In this article, the following optimal
is shown in equation (3) targets were selected:
(a) The maximum optimal target of separa-
ðymax  ymin Þ 3 ðx  xmin Þ tion efficiency
y= + ymin ð3Þ
ðxmax  xmin Þ
min f1 = minð100  hÞ ð5Þ
where x is the value before normalization and y is the
normalized value; xmin, xmax, ymin, and ymax represent (b) The minimum optimal target of the pres-
the minimum value before normalization, the maxi- sure drop
mum value before normalization, the normalized mini-
minf2 = minðDPÞ ð6Þ
mum value, and the normalized maximum value,
respectively.
where h is the predicted separation efficiency of BP
The neural network estimation result with
neural network (%) and DP is the predicted pressure
Levenberg–Marquardt algorithm is accurate and the
drop of BP neural network (Pa).
stability of the network is better.21 Therefore, the
3. Restrictions
trainlm function is selected as a training function to
update weight and bias values. Max_fail indicates max- 20 ł D ł 40
imum number of validation checks before training is
stopped and the value is 20. Both the hidden layer acti- 90 ł a ł 130
vation function and the output layer activation func- 160 ł H ł 200
tion are tansig, and the training function is trainlm. 8 ł L ł 16
When training the BP neural network, performance 2łvł6
function is mse, the maximum iteration number is
90 ł T ł 130
1000, and the target error is 10–5. Other parameters are
default values. 4. Settings of parameters

The population size is 100, the maximum number of

Multi-target optimization of the genetic algorithm iterations the genetic algorithm performs is 500, the
A genetic algorithm is a global probability searching Pareto front population fraction is 0.3, the crossover
method that simulates genetic selection and natural fraction is 0.8, and the function tolerance in the stop-
elimination. Its main characteristic is that the searching ping criteria is 10–4.
does not rely on the gradient information or require
differentiation of the solution function. Its parallel and Results and discussion
global searching abilities are suitable for a large-scale
non-linear optimization model. Therefore, this method Results analysis of orthogonal experimental design
has been applied to solving various optimization The simulation results of the orthogonal experiment,
issues.22,23 In this article, we used a genetic algorithm listed in Table 3, show that experiment E11 has the
for multi-target optimization of the predictive model highest efficiency and the lowest pressure drop. It is
established by BP neural network. The multi-target the optimal solution of the orthogonal experiment. The
optimization mathematic model is as follows: average efficiency and average pressure drop of 25
orthogonal experiments were 87.66% and 50.68 Pa,
1. Optimization variables. In this article, the opti- respectively.
mal model uses six design variables, which are Range analysis was used to analyze the simulation
the vane spacing D (mm), vane angle a (o), vane results from orthogonal experiments. The average and
height H (mm), drainage channel length L range of h at different levels of various factors are
(mm), gas velocity v (m/s), and temperature T shown in Table 4. The average and range of DP at dif-
(°C). Different values of these variables mean ferent levels of various factors are shown in Table 5.
different solutions, as shown in equation (4) In Tables 4 and 5, a larger range indicates a greater
difference in separation efficiency and pressure drop at
x = ðD, a, H, L, v, T ÞT ð4Þ different levels for the factor. It also indicates that the
factor also has a greater impact on the separation effi-
2. Objective function. The objective function is a ciency and pressure drop, which is the main parameter,
mathematical expression of an optimization and vice versa. Table 4 shows that the order of impact
goal expressed by optimization variables and is of these factors on separation efficiency h is
an evaluation criterion for the quality of the a . v . D . T . H . L. Table 5 shows that the order
He et al. 7

Table 3. The Fluent simulation results of the orthogonal experiment.

Trial no. D A H L v T h (%) DP (Pa)

E1 20 90 160 8 2 90 100.00 39.82

E2 20 100 170 10 3 100 100.00 53.49
E3 20 110 180 12 4 110 100.00 56.12
E4 20 120 190 14 5 120 100.00 52.50
E5 20 130 200 16 6 130 100.00 44.04
E6 25 90 170 12 5 130 100.00 128.85
E7 25 100 180 14 6 90 100.00 120.83
E8 25 110 190 16 2 100 96.66 9.81
E9 25 120 200 8 3 110 98.92 16.14
E10 25 130 160 10 4 120 97.53 18.65
E11 30 90 180 16 3 120 100.00 36.91
E12 30 100 190 8 4 130 100.00 55.12
E13 30 110 200 10 5 90 100.00 57.74
E14 30 120 160 12 6 100 100.00 63.97
E15 30 130 170 14 2 110 30.40 3.60
E16 35 90 190 10 6 110 100.00 147.40
E17 35 100 200 12 2 120 85.07 10.74
E18 35 110 160 14 3 130 82.95 17.91
E19 35 120 170 16 4 90 62.00 20.79
E20 35 130 180 8 5 100 75.77 39.71
E21 40 90 200 14 4 100 99.99 55.30
E22 40 100 160 16 5 110 99.87 69.18
E23 40 110 170 8 6 120 100.00 132.76
E24 40 120 180 10 2 130 33.89 7.03
E25 40 130 190 12 3 90 28.48 8.71

Table 4. The average and range of h at different levels of various factors.

Level Average of separation efficiency (%)

D A H L v T

1 100.00 100.00 96.07 94.94 69.20 78.10

2 98.62 96.99 78.48 86.28 82.07 94.48
3 86.08 95.92 81.93 82.71 91.90 85.84
4 81.16 78.96 85.03 82.67 95.13 96.52
5 72.45 66.44 96.80 91.71 100.00 83.37
Range 27.55 33.56 18.32 12.27 30.80 18.42

Table 5. The average and range of DP at different levels of various factors.

Level Average of pressure drop (Pa)

D a H L v T

1 49.19 81.66 41.91 56.71 14.20 49.58

2 58.86 61.87 67.90 56.86 26.63 44.46
3 43.47 54.87 52.12 53.68 41.20 58.49
4 47.31 32.09 54.71 50.03 69.60 50.31
5 54.60 22.94 36.79 36.15 101.80 50.59
Range 15.39 58.71 31.11 20.72 87.60 14.03

of impact of these factors on the pressure drop DP is gas velocity v are the two main factors that have the
v . a . H . L . D . T. Therefore, vane angle a and most significant impact on h and DP.
8 Advances in Mechanical Engineering

Figure 5. Correlation coefficient R: (a) training data set, (b) validation data set, (c) test data set, and (d) overall data set.

The prediction performance of the BP neural network

model
Figure 5(a)–(d) shows the correlation coefficient R of
the expected output and the network prediction out-
puts from the training data set, validation data set, test
data set, and overall data set. In the regression fitting,
a correlation coefficient closer to 1 indicates a higher
accuracy model.24 Figure 5 shows that the R values for
the training data set, validation data set, test data set,
and overall data set are 0.99968, 0.99971. 0.99918, and
0.99961, respectively, which show the excellent consis-
tency between the expected value and the value pre-
dicted. Hence, the model established by BP neural
network can be used to predict the separation efficiency
and pressure drop.

Results of multi-target optimization genetic algorithm Figure 6. Pareto front.

Using the genetic algorithm to optimize the predictive
model established by BP neural network, 30 solutions
are obtained, as shown in Figure 6. The average effi- two solutions are almost completely equal, a fork point
ciency and average pressure drop of the 30 sets of solu- in Figure 6 is used to highlight the two coincident data
tions were 100% and 49.07 Pa, respectively. Because points.
He et al. 9

Table 6. Multi-target optimization results.

Trial no. D a H L v T h (%) DP (Pa)

G1 22 99 187 14 2.37 127 100.00 20.75

G2 21 100 187 14 2.81 127 100.00 34.82
G3 21 100 187 14 3.06 127 100.00 44.28
G4 21 97 188 15 3.26 129 100.00 64.22
G5 20 97 187 15 3.51 130 100.00 79.40
G6 20 98 187 15 3.42 128 100.00 89.83
G7 20 98 188 15 3.86 129 100.00 115.85
G8 20 96 187 15 3.85 130 100.00 122.33
G9 20 97 187 15 3.95 129 100.00 137.33
G10 20 96 187 14 4.12 129 100.00 153.61
G11 20 97 187 15 4.32 129 100.00 170.54
G12 20 96 187 15 4.36 130 100.00 189.88
G13 20 96 187 15 4.62 130 100.00 217.92

Table 7. Fluent simulation results of five groups of optimal

parameters combination.

Trial no. G1 G2 G3 G4 G5

h (%) 100 100 100 100 100

DP (Pa) 27.4 41.53 48.84 64.21 85.31

Because solutions obtained by multi-objective opti-

mization of genetic algorithm are not integers, and the
size of the demister produced by factories is not so pre-
cise, we simplify the solution in Figure 7. The vane
spacing, vane height, vane angle, drainage channel
length, and temperature are rounded to integer number,
and the gas velocity, f1, and f2 are all accurate to two
Figure 7. Comparison of the model prediction value and
decimal places. In this article, the separation efficiency
simulation value.
is more important than the pressure drop of the vane
demister. Therefore, 13 sets of solutions with f1 equal
to 0 are selected from the simplified 30 sets of solutions,
process. Comparison of the Fluent simulation results
and the original solutions of G1 to G13 before the sim-
and the model prediction results is shown in Figure 7.
plification are shown in Table 9 (refer Appendix 2).
The average absolute deviation (AAD) of all para-
Then, f1 is converted to h and 13 sets of solutions are
meters was calculated using equation (8). Figure 7
arranged in ascending order of pressure drop to obtain
shows that the model prediction results and Fluent
Table 6.
simulation results have a small deviation, where the
In Table 6, G1 is the optimal solution with 100%
AAD of separation efficiency (AADh) is 0 and the
efficiency and the lowest pressure drop. The results
AAD of the pressure drop (AADDP) is 4.768 Pa.
from genetic algorithm are derived from the predicted
Therefore, through the validation of simulations in
values of the BP neural network model, and there may
Fluent, the model established by BP neural network
be some errors between the predicted values of the
also shows a high accuracy of prediction on non-
model and the actual CFD simulation results. To fur-
sample data
ther verify the prediction accuracy of the model and
obtain the true optimal solution, the five sets of data
with 100% efficiency and the lowest pressure drop in 1X N
AAD = jPi  Ci j ð8Þ
Table 6, namely, G1, G2, G3, G4, and G5, were N i=1
selected for CFD simulation, and the results are shown
in Table 7. where N is the number of data sets, Pi is the prediction
Table 7 shows that G1 has the lowest pressure drop, value of the model, and Ci is the calculated value of the
which is the optimal solution of this optimization model.
10 Advances in Mechanical Engineering

Table 8. Comparison of optimal results. gas velocity are the two main factors that affect
the separation efficiency and pressure drop.
Trial D A H L v T h (%) DP (Pa) 2. Using the Fluent simulation results as a sample
no.
database, the prediction model for the separa-
E11 30 90 180 16 3 120 100 36.91 tion efficiency and pressure drop was created
G1 22 99 184 14 2.37 127 100 27.4 based on a BP neural network. Using a genetic
algorithm for optimization, the optimal separa-
tion efficiency is 100% and the optimal pressure
drop is 27.4 Pa. Comparison of five groups of
Discussion optimal results with Fluent simulation results
In Table 8, E11 is the optimal solution of orthogonal shows that AADh is 0 and AADDP is 4.768 Pa.
design and G1 is the optimal solution of genetic algo- 3. The optimal pressure drop from the genetic
rithm. The separation efficiency of both solutions is algorithm is 25.77% lower than that from
100%. However, the average separation efficiency of orthogonal design, which indicates that the
the genetic algorithm is higher and the pressure drop of genetic algorithm results are better than an
G1 is 25.77% lower than that of E11, which indicates orthogonal design of experimental results.
that the genetic algorithm has a higher optimization Hence, we propose the combined method of a
performance. On one hand, from the calculation pro- neural network and intelligent algorithm,
cess, the orthogonal design discretizes the six continu- which shows the high feasibility of the vane
ous variables studied, and the results are also non- demister optimization design and provides a
continuous, which shows no continuous prediction new idea for optimization design of the vane
ability. The optimal solution from the orthogonal demister.
design is one of the sample data, and non-sample data
cannot be obtained. On the other hand, the prediction
Declaration of conflicting interests
model of the BP neural network established on limited
samples can predict continuously and shows an excel- The author(s) declared no potential conflicts of interest with
lent non-linear prediction ability. As a global optimal respect to the research, authorship, and/or publication of this
article.
algorithm, a genetic algorithm can search within the
feasible domain and obtain the parameter combination
of non-sample data, which can further generate a better Funding
optimization effect. Even though the orthogonal design The author(s) disclosed receipt of the following financial sup-
results are worse than those of the genetic algorithm, it port for the research, authorship, and/or publication of this
can generate feasible parameter combinations from less article: This paper is funded by the State Key Laboratory of
calculation. Oil and Gas Reservoir Geology and Exploitation in
Overall, the prediction model of the separation effi- Southwest Petroleum University (The funding number is
ciency and pressure drop was established using the BP 448). The authors express their gratitude to the funding.
neural network, and using the genetic algorithm as the
optimal method shows the high feasibility of vane dem- ORCID iDs
ister optimization design. Therefore, when different San He https://orcid.org/0000-0002-3306-3951
optimal parameters are considered for other types of Hang Liu https://orcid.org/0000-0001-8347-4313
demisters, the optimization idea of this article can also
be used. First prepare sample data, then use a neural
network to establish a prediction model of demist per- References
formance parameters, and finally use intelligent algo- 1. Venkatesan G, Kulasekharan N and Iniyan S. Numerical
rithm to optimize the model according to different analysis of curved vane demisters in estimating water
process requirements. droplet separation efficiency. Desalination 2014; 339:
40–53.
2. Kong Q. The mechanical performance of the wave-plate
Conclusion mist eliminator. Master’s Thesis, Tianjin University,
Tianjin, China, 2013.
1. An orthogonal array of 25 cases (L25) for six
3. Estakhrsar MHH and Rafee R. Effects of wavelength
parameters with five levels was selected. The and number of bends on the performance of zigzag dem-
optimal parameter combination was determined isters with drainage channels. Appl Math Modell 2016;
according to the experimental results, where the 40: 685–699.
separation efficiency is 100% and the pressure 4. Xu SJ, Yao Z and Zhu YY. Numerical simulation on
drop is 36.91 Pa. The range analysis of orthogo- two phase flow in demister with corrugated baffle. J Univ
nal design results show that the vane angle and Shanghai Sci Technol 2007; 29: 275–280.
He et al. 11

5. Wang X. Experimental study and numerical simulation 21. Song L, Lin GU, Dong H, et al. Depth estimation of
of characteristics of demister of absorbing tower in desul- facial feature points based on optimized BP neural net-
furization system. Chin J Environ Eng 2008; 2: work. J Xian Polytech Univ 2017; 31: 551–555.
1529–1534. 22. Mitra AC, Desai GJ, Patwardhan SR, et al. Optimization
6. Wang YI and James PW. The calculation of wave-plate of passive vehicle suspension system by genetic algorithm.
demister efficiencies using numerical simulation of the Proced Eng 2016; 144: 1158–1166.
flow field and droplet motion. Chem Eng Res Des 1998; 23. Moeini A and Zargarabadi MR. Genetic algorithm opti-
76: 980–985. mization of film cooling effectiveness over a rotating
7. Venkatesan G, Kulasekharan N and Iniyan S. Influence blade. Int J Therm Sci 2018; 125: 248–255.
of turbulence models on the performance prediction of 24. Yin H, Wang K, Zhang T, et al. Fault prediction based
flow through curved vane demisters. Desalination 2013; on PSO-BP neural network about wheel and axle box of
329: 19–28. bogie in urban rail train. Complex Syst Complex Sci 2015;
8. Zhao J, Jin B and Zhong Z. Study of the separation effi- 12: 97–103.
ciency of a demister vane with response surface metho-
dology. J Hazard Mater 2007; 147: 363–369.
9. Narimani E and Shahhoseini S. Optimization of vane Appendix 1
mist eliminators. Appl Therm Eng 2011; 31: 188–193.
10. Venkatesan G, Kulasekharan N and Iniyan S. Design
Notation
and selection of curved vane demisters using Taguchi AADh average absolute deviation of separation
based CFD analysis. Desalination 2014; 354: 39–52. efficiency (%)
11. Venkatesan G, Kulasekharan N, Muthukumar V, et al.
AADDP average absolute deviation of pressure
Regression analysis of a curved vane demister with Tagu-
drop (Pa)
chi based optimization. Desalination 2015; 370: 33–43.
12. Galletti C, Brunazzi E and Tognotti L. A numerical
Ci calculated value of simulation in Fluent
model for gas flow and droplet motion in wave-plate mist software
eliminators with drainage channels. Chem Eng Sci 2008; D vane spacing (mm)
63: 5639–5652. f1 optimal target of genetic algorithm,
13. Rafee R, Rahimzadeh H and Ahmadi G. Numerical difference between 100 and h
simulations of airflow and droplet transport in a wave- f2 optimal target of genetic algorithm,
plate mist eliminator. Chem Eng Res Des 2010; 88: pressure drop
1393–1404. H vane height (mm)
14. Xie H. Study on the performance of normal distribution arc L length of drainage channel (mm)
plate in demister in the gas water mist transportation sys- m number of droplets at the inlet of the
tem. Master’s Thesis, Henan Polytechnic University, Jiao- demister
zuo, China, 2014.
n number of droplets at the outlet of the
15. Cheng C, Cheng X, Dai N, et al. Prediction of facial
demister
deformation after complete denture prosthesis using BP
neural network. Comput Biol Med 2015; 66: 103–112. N number of data sets
16. Ma D, Zhou T, Chen J, et al. Supercritical water heat Nin number of neurons in the inlet layer
transfer coefficient prediction analysis based on BP neural Nhid number of neurons in the hidden layer
network. Nucl Eng Des 2017; 320: 400–408. Pi prediction value of neural network
17. Shen B. The study of stock price prediction based on BP R correlation coefficient
neural network. Master’s Thesis, Hunan University, T temperature (°C)
Changsha, China, 2010. v gas velocity (m/s)
18. Jiao B and Ye M. Determination of hidden unit number x value before normalization
in a BP neural network. J Shanghai Dianji Univ 2013; 16: xmin minimum value before normalization
113–116. xmax maximum value before normalization
19. Fan Q. Research on micro-vortex coagulation dosing con-
y normalized value
trol model based on genetic algorithm and BP neural net-
ymin normalized minimum value
work. Master’s Thesis, East China Jiaotong University,
Nanchang, China, 2018.
ymax normalized maximum value
20. Kou F. Short-term traffic flow prediction based on adap- a vane angle (°)
tive artificial fish swarm algorithm BP recurrent neural DP pressure drop (Pa)
network. Master’s Thesis, Beijing Jiaotong University, h separation efficiency (%)
Beijing, China, 2018.
12 Advances in Mechanical Engineering

Appendix 2

Table 9. Original multi-objective solutions of G1–G13.

Trial no. f1 f2

G1 4.98677E205 20.74858478
G2 8.14972E206 34.81698004
G3 4.2129E206 44.2825304
G4 3.36303E207 64.21559021
G5 2.04827E207 79.39617224
G6 1.8287E207 89.83077928
G7 7.68879E208 115.848598
G8 6.56642E208 122.3291917
G9 6.34791E208 137.3267092
G10 5.24537E208 153.6112148
G11 4.38243E208 170.5428354
G12 3.58288E208 189.8839979
G13 3.3879E208 217.9190998