Professional Documents
Culture Documents
Jert 20 1313 2
Jert 20 1313 2
Jert 20 1313 2
net/publication/342402119
Article in Journal of Energy Resources Technology, Transactions of the ASME · June 2020
DOI: 10.1115/1.4047593
CITATIONS READS
105 1,719
4 authors, including:
All content following this page was uploaded by M. S. Shadloo on 17 July 2020.
Estimation of pressure drop of two-phase flow in horizontal long pipes using artificial
neural networks
d
te
2
Institute of Chemical Process Engineering, University of Stuttgart, Stuttgart, 70199, Germany
di
3
School of Chemical Engineering, University of Birmingham, B15 2TT, United Kingdom
e
4
Sustainable Management of Natural Resources and Environment Research Group, Faculty of Environment
py
and Labour Safety, Ton Duc Thang University, Ho Chi Minh City, Vietnam, arashkarimipour@tdtu.edu.vn
5
Fluid Mechanics, Thermal Engineering and Multiphase Flow Research Lab. (FUTURE), Department of
Co
Mechanical Engineering, Faculty of Engineering, King Mongkut’s University of Technology Thonburi,
Bangmod, Bangkok 10140, Thailand
ot
6
The Academy of Science, The Royal Society of Thailand, Sanam Suea Pa, Dusit, Bangkok 10300, Thailand
tN
Abstract
rip
sc
Gas-liquid two-phase flows through long pipelines are one of the most common cases found in
nu
chemical, oil, and gas industries. In contrast to the gas/Newtonian liquid systems, the pressure drop
has rarely been investigated for two-phase gas/non-Newtonian liquid systems in pipe flows. In this
Ma
regard, an Artificial Neural Networks (ANN) model is presented by employing a large number of
experimental data to predict the pressure drop for a wide range of operating conditions, pipe
ed
(MLPNN) model, the predicted pressure drop is in a good agreement with the experimental results.
In most cases, the deviation of the predicted pressure drop from the experimental data does not
ce
exceed 5%. It is observed that the MLPNN provides more accurate results for horizontal pipelines
Ac
in comparison with other empirical correlations that are commonly used in industrial applications.
Keywords: Two-phase flow, Gas/non-Newtonian liquid, Pressure drop, Artificial neural networks
(ANN)
*Corresponding Author
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
1. Introduction
d
te
Over the past few decades, some experimental and theoretical investigations on gas/non-
di
Newtonian fluid flows have resulted in the development of some correlations for prediction of
e
pressure drop in pipelines. There are several principle studies that present the data of flow patterns
py
in vertical and horizontal pipelines for gas/non-Newtonian liquid mixtures [5–9]. Empirical
correlations are usually applicable only in limited ranges of fluid characteristics; hence, special
Co
attention should be paid when extrapolating outside the experimental conditions range.
ot
In the past few years, the co-current flow of two-phase gas-liquid systems has been extensively
tN
investigated in inclined and horizontal pipelines. Theoretical and experimental studies were carried
out to determine crucial hydrodynamic characteristics of gas-liquid systems such as void fraction,
rip
pressure drop, and multiphase flow patterns. In order to accurately estimate the pressure drop and
void fraction at a given flow condition, it is crucial to identify the flow pattern of such multiphase
sc
systems [10–15]. On the other hand, identification of flow patterns, and the calculation of void
nu
fraction and pressure drop are mainly available for Newtonian liquids since numerous
experimental studies are performed on two-phase gas/Newtonian systems in pipe flows [16–21].
Ma
There are several models available in the literature for predicting the pressure drop of gas/non-
ed
Newtonian fluids in pipelines. Richardson and Chhabra [22] modified the map of the flow pattern
proposed by Mandhane et al.[23] for horizontal pipes, to represent the flow pattern of gas/non-
pt
Newtonian systems based on existing data for gas/shear-thinning liquids mixtures. Dziubinski [24]
ce
employed the concept of loss coefficient to establish an expression for pressure drop during a two-
Ac
phase gas/non-Newtonian liquid intermittent flow. Ruiz-Viera et al.[25] also examined the
air/lubricating grease mixture flow in various geometries with coarse or smooth surfaces that
resulted in an experimental model for calculation of pressure drop in two-phase flows. These
models can be easily solved for a pressure drop of two-phase flows since they do not require much
information from the flow structures.
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
Heywood and Charles [26] modified a correlation that was initially proposed by Taitel and Dukler
[27] for gas/liquid Newtonian flow for predicting the pressure drop and liquid holdup in a unified
stratified flow. Eisenberg and Weinberger [28] studied horizontal pipes containing gas/power-law
liquids for the annular flow regime. It should be noted that there are limited number of methods
d
Prediction of the flow pattern is the primary stage in almost all studies. A comparison of these
te
models indicated some inconsistencies in their performance at different flow conditions. Hence, it
di
is crucial to choose the most suitable correlation for the specific flow. In cases that simple
e
mathematical models are unable to describe complex problems, the artificial neural network
py
(ANN) technique can act as a powerful tool to solve these problems. One of the advantages of the
Co
ANN techniques is that there is no need to have a deep understanding of the physical phenomena.
The exact prediction of flow pattern, liquid holdup, and pressure drop is complex for multiphase
ot
flows in pipelines, especially at inclined conditions. So, many researchers have utilized the ANN
tN
algorithms for such complex gas-liquid systems [29–31]. For instance, Cai et al. [32] identified
the air-water flow regimes in a horizontal pipe using Kohonen self-organizing neural network. A
rip
new model was also proposed by Osman and Aggour [33] for the identification of different flow
regimes and the prediction of holdup in gas-liquid flow in horizontal pipes employing the three-
sc
layer back-propagation neural networks. The liquid holdup was also predicted by another ANN
nu
model proposed by Shippen and Scott for a two-phase horizontal flow [34].
Ma
In this paper, we aim to investigate the two-phase flow characteristics, namely the pressure drop
in pipelines for gas/non-Newtonian liquid flow mixtures utilizing an ANN method. A comparison
ed
was made between the experimental pressure drop and those obtained from the ANN method. The
flow behavior index, pipe diameter, consistency coefficient, the density of the liquid, liquid
pt
velocity, and gas velocity were selected as the inputs of ANN. This follows by a comparison of
ce
2. Methods
Dziubinski [24] suggested an empirical correlation for single-phase flow in horizontal pipes for
predicting the pressure drop at intermittent flow regimes:
d
884? 𝐾 ⎥
( ) = ( )) ⎢ ⎥ 𝜆)
𝑑𝑙 &' 𝑑𝑙 ⎢ ?.;BC
te
8𝑢 ;48
45
𝐷 ) 𝜌) ⎥
⎣ 1 + 1.036 × 10 6 884? 𝐾 A ⎦
e di
H'
where , 𝐷, 𝜆) , 𝑛, and 𝐾 are pressure drop, pipe diameter, input liquid volume fraction, flow
py
H)
behavior index, and fluid consistency coefficient, respectively. The mixture velocity is equal to
Co
𝑢: = 𝑢K) + 𝑢KL , and 𝑢KL and 𝑢K) are the superficial gas and liquid velocities, respectively. 𝜌: is
the average density of the slug unit calculated by 𝜌: = (1 − 𝜆) )𝜌L + 𝜆) 𝜌) . Moreover, 𝑡𝑝 and l
ot
subscript denote two-phase and liquid. tN
The average error of this correlation is ± 15 % and Dziubinski used regression to reach this
rip
equation. It should be noted here that this model is empirical and cannot be extrapolated without
taking special care outside the experimental condition range [35].
sc
Taitel and Barnea [36] have extensively investigated the mechanical model applied to the slug
Ma
flow of gas/Newtonian fluid in a steady-state condition. The slug units might consist of two distinct
patterns i.e., liquid and gas slugs as depicted in Figure 1.
ed
pt
ce
Ac
If it is assumed that the liquid film is free of gas bubbles and both fluid phases are incompressible,
the total pressure drop is equal to the summation of the pressure drops for the gas and liquid slugs
as follows:
d
In this equation, the 𝑔𝑠 and 𝑙𝑠 subscripts denote the gas and liquid slug, respectively. This two-
te
phase model has been recently modified by Xu et al. [37] to become applicable to power-law
di
fluids. Several physical variables of gas/Newtonian fluid systems were included in the gas/non-
e
Newtonian liquid flow model. The results for carboxymethyl cellulose (CMC) solutions indicate
py
that the two-fluid model complies well with the pressure drop and reduction in drag data.
Co
ot
2.1.3 Developed Model tN
If the reduction of pressure across the gas slug is neglected, the homogeneous model can be
utilized, at the same velocity, for pressure drop calculation along the liquid slug.
rip
&' )K W
nu
Here, 𝑢: = 𝑢K) + 𝑢KL is the superficial velocity of gas and liquid mixture. The term 𝜌: =
(1 − 𝑎K )𝜌L + 𝑎K 𝜌) is the liquid slug average density. 𝑎K is the holdup of liquid slug, 𝜅 denotes the
Ma
liquid slug zone length to the slug unit length ratio and f is the friction factor. If the flow model is
assumed an idealized intermittent one, 𝜅 ≈ 𝜆) would be calculated [35].
ed
Xu [38] recommended that the liquid slug holdup of the gas/non-Newtonian mixture flow can be
pt
expressed as:
ce
(?4K[8\)].]^ (4)
𝑎K = ?_B.?``×?ab^cd g.hh^ 0 ≤ 𝜃 ≤ 75
Ac
ef
In this equation, 𝜃 is the inclination angle from the horizontal and Re is the Reynolds number.
ANNs are a powerful tool in investigations of neurobiology and human information processing by
cognitive science. It stimulates the neurons in the network structure and characteristics to achieve
various complex information processing functions. During the ANN learning process, it is feasible
to found a functional relationship between two data spaces and then replicate/generalize them
d
characteristics of neurons in a simulated network structure for this purpose. The learning stage
te
allows the ANN to create a functional relationship between two datasets by which the ANN will
di
utilize them during a recall process [39–42].
e
The ANN is a mathematical system composed of neurons as simple processing elements that are
py
generated in single or multilayers. All of the ANN models were formed from a relatively large
Co
number of interconnected key elements named artificial neurons [24]. The ANNs can be formed
by placing a specified number of neurons in some layers. The information from an external source
ot
is received by neurons that are located in the first layer. The processed information will be
tN
transfered to the next layer(s) after undergoing some mathematical operations. The performed
mathematical operations on the input data 𝑥[ are explained in Equation. 5. The output of the
rip
𝑦[ = 𝑔 op 𝑤[ 𝑥[ + 𝑏t
[r?
nu
Here, 𝑤[ denotes the weight associated with the connection between the entry signal and 𝑖 th
neuron. The bias and activation function of the neuron are 𝑏 and 𝑔, respectively.
Ma
The most commonly used ANNs are multilayer perceptron neural networks (MLPNNs) which is
a class of artificial neural networks in which the neurons use a nonlinear activation function for
ed
learning [43,44]. Although the different architecture of ANNs, including radial basis function
pt
neural network (RBFNN), cascade feed-forward neural network (CFNN), generalized regression
ce
neural network (GRNN), and pattern recognition neural network (PRNN) can be used to make
predictions based on data [45–47].
Ac
A trial and error method was adopted with some error indices minimizing functions for the
determination of the number of hidden neurons. In this work, mean square error (MSE), average
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
absolute relative deviation (AARD), and root mean square error (RMSE) were minimized and the
regression coefficient (R2) was maximized as the values of training and validation datasets for
determining the number of hidden neurons. The MSE, AARD, R2, and RMSE are defined as:
d
i =1
te
N æ Yi ,act - Yi , pred ö (7)
1
AARD % = å ç
ç
÷ ´ 100
÷
di
N i =1 è Yi ,act ø
e
(8)
- Yact ) - å (Yi ,act - Yi , pred )
N N
å (Y
2 2
py
i ,act
R2 = i =1
N
i =1
å (Y - Yact )
2
Co
i ,act
i =1
2 (9)
å (Y - Yi , pred )
N
1
ot
RMSE = i ,act
N i =1
tN
where N stands for the number of data, and 𝑌wxy& denotes the mean of actual values. 𝑌[.xy& and
rip
Table 1 presents the properties of the experimental pressure drop databank which were obtained
ed
from the literature along with their pertinent references. A set of 511 data points were used for
pt
evaluation of the suggested model and three other associated empirical equations and the
prediction of the pressure drop in horizontal pipes containing gas/ non-Newtonian liquid.
ce
0.0417–0.207 0.103 18- 48.5 1265 -1310 0.28–1.35 0–3.1 Kaolin [49]
d
te
By the literary survey, it was concluded that many parameters including properties of the liquid,
di
preparation procedure, and the operational conditions can affect the pressure drop of two-phase
e
flow [50–53]. Therefore, all variables including the pipe diameter, flow behavior index,
py
consistency coefficient, the density of the liquid, liquid velocity, and gas velocity were selected as
Co
the inputs (denoted by input1, input2, input3, input4, input5. and input6, respectively) of our
intelligent paradigms. It should be noted that the pressure drop was considered as the response
ot
variable. tN
rip
Analysis of a correlation matrix is used for determining how an individual independent variable
sc
can affect the dependent variable [54]. We used Pearson’s correlation to evaluate the strength of a
nu
linear relationship between the dependent-independent pairs of variables in our system. The power
of the linear relationship is calculated by the division of covariance by the multiplication of
Ma
variables standard deviations. The value of the results varies in the range of -1 to +1; where +1 is
an indication that the linear relation is positive, -1 indicates a negative linear relation, and 0
ed
indicates that there is no relation between them. The results of the linear relationship between the
pressure drop and various independent variables are presented in Figure. 2.
pt
ce
Ac
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
0.8
0.6
d
0.2
te
Input1 Input3
di
0
Input2 Input4 Input5 Input6
e
-0.2
py
-0.4
Co
-0.6
Independent variables
ot
tN
Figure. 2. The linear relationship between different sets of independent-dependent variables.
rip
It can be seen that the most powerful direct and indirect relations belong to flow behavior index,
liquid velocity and fluid consistency coefficient, respectively and pipe diameter has the least
sc
It should be noted that to enhance the rate of convergence in the training stage and also prevent
ed
the ANN model parameters from saturation, Equation 10 was used to map the experimental data
pt
V - Vmin
V = 0.98 ´ + 0.01 (10)
Vmax - Vmin
Ac
Where 𝑉 denotes any dependent or independent variable, 𝑉:x| is the maximum and 𝑉:[8 is the
minimum value of each variable, and the normalized values are indicated by 𝑉w .
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
The collected data points were randomly divided into two groups of training and testing points.
85% of these data were used as the training dataset and were used for adjusting the weights and
d
te
A trial and error procedure is normally adopted to identify the optimum structure of ANNs. The
di
optimum configuration is determined by altering the number of layers and the number of included
e
neurons in each one. According to this method, with the function of minimizing some error indices,
py
it was adopted for the determination of the number of hidden neurons. In this works, RMSE,
AARD, and MSE were minimized and R2 was maximized as the values of training and validation
Co
datasets for determination of the number of hidden neurons.
It has been affirmed that any multivariable function can be precisely correlated by an MLP network
ot
that has just one hidden layer. Therefore, in this study, we used an MLP network with a single
tN
layer to predict the pressure drop.
Thus, the number of MLPNN (one hidden layer) hidden neurons with one dependent variable
rip
𝑁 ≤ 10
Ma
The number of hidden neurons could change between the lowest and highest quantity (1 -10), and
each network is trained and then tested fifty times. Table 2 presents the AARD%, RMSE, R2, and
ed
MSE values obtained during testing, training, and total datasets for different the number of hidden
pt
neurons of MLPNN.
ce
Ac
Table 2. Sensitivity analysis results to identify the best hidden neuron of MLPNN
Statistical index
Hidden neuron dataset
AARD% MSE RMSE R2
Train 26.67 0.1503 0.3876 0.54770
1 Test 26.36 0.2154 0.4641 0.57351
Total 26.63 0.1601 0.4001 0.55029
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
d
Total 10.47 0.0078 0.0882 0.98249
te
Train 8.13 0.0061 0.0778 0.98736
5 Test 11.92 0.0224 0.1496 0.96188
di
Total 8.70 0.0085 0.0923 0.98278
e
Train 6.01 0.0029 0.0538 0.99313
6 Test 7.88 0.0262 0.1617 0.95618
py
Total 6.29 0.0064 0.0800 0.98567
Train 5.05 0.0022 0.0472 0.99500
Co
7 Test 6.74 0.0037 0.0607 0.99153
Total 5.30 0.0024 0.0495 0.99448
Train 4.90 0.0024 0.0493 0.99501
ot
8 Test 7.63 0.0073 0.0852 0.96685
Total 5.32 0.0032 0.0562 0.99287
tN
Train 4.64 0.0019 0.0433 0.99588
9 Test 6.22 0.0037 0.0609 0.99036
rip
It shows that AARD%, MSE, RMSE, and R2 should take the values of 4.58, 0.0025, 0.05, and
0.99438, respectively to have the optimum MLP network to predict total experimental data points.
The values obtained for error indices and statistical measures verify that pressure drop can be
ed
Similar to the MLPNN, for other models including CFNN, GRNN, and RBFNN, we considered
ce
the fact that the training experiments should be at least 5 times greater than the number of the
Ac
network parameters. Accordingly, the optimum number of hidden neurons for each model was
determined.
When the optimal number of hidden neurons and the most appropriate training algorithm was
selected for the MLPNN, it would be preferable to compare the accuracy of its prediction with
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
other models like CFNN, GRNN, and RBFNN. A comparison between the predictive capabilities
of the MLP neural network model with CFNN, GRNN, and RBFNN is provided in Table 3.
d
MLPNN Test 5.83 0.0064 0.0803 0.98622
te
Total 4.58 0.0025 0.0500 0.99438
Train 4.69 0.0022 0.0474 0.99486
di
CFNN Test 5.75 0.0039 0.0622 0.99310
Total 4.85 0.0025 0.0499 0.99449
e
Train 0.49 0.0002 0.0124 0.99964
py
GRNN Test 9.62 0.0122 0.1103 0.97801
Total 1.86 0.0020 0.0443 0.99557
Train 19.44 0.0595 0.2439 0.86282
Co
RBFNN Test 23.15 0.0591 0.2432 0.83791
Total 20.00 0.0594 0.2438 0.85772
ot
The number of hidden neurons in all ANN models (CFNN, GRNN, RBFNN, and MLPNN) was
tN
considered at the optimum number to enable a fair comparison between them.
rip
The superiority of the MLPNN network was proved by the results of pressure drop prediction. The
most inaccurate prediction belongs to the RBFNN and CFNN models. It must be pointed that
sc
though the GRNN is more accurate in terms of statistical indices, however, the accuracy of the
nu
training phase is about 19 times higher than the accuracy of the experimental phase, this model
has probably over fitted and cannot be introduced as the best model.
Ma
Amongst 2060 intelligent models (500 MLPNN, 450 CFNN, 110 GRNN, and 1000 RBFNN) this
model was chosen based on the four sensitivity indices, including MSE, RMSE, AARD%, and R2.
ed
Figure 3 clearly shows that the MLPNN with ten hidden neurons come to the MSE value of
pt
4.03 × 1045 after 625 times adjustment of bias and weight by the training stage. Subsequently,
ce
there was no improvement in the results and the network training was practically terminated.
Ac
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
MSE
Desired MSE
d
te
e di
-3
10
py
Co
0 100 200 300 400 500 600
Number of adjustemnts of ANN model
ot
Figure 3. Variation of MSE of the optimum MLPNN during the training stage.
tN
It is evident that the optimum structure (6-10-1 structure) is an MLP model having a single hidden
rip
layer that contains ten hidden neurons (bold rows). This model has a 𝑡𝑎𝑛𝑠𝑖𝑔 transfer function
which is located in the hidden layer and the 𝑙𝑜𝑔𝑠𝑖𝑔 transfer function is placed in the output layer,
sc
The amazing ability of the suggested MLPNN in pressure drop estimation is indicated in Figure.
4 in a testing (blue area) and training (red area) datasets. The predicted values for pressure drop
Ma
perfectly match their related real data points; it means that the MLPNN model can perfectly predict
the pressure drop in these conditions.
ed
pt
ce
Ac
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
3 Training Dataset
Test Dataset
Equality Line
2.5
d
te
1.5
di
1
e
py
0.5
Co
0.5 1 1.5 2 2.5 3
Experimental
ot
Figure 4. The plot of experimental values versus predicted values by MLPNN
tN
rip
Figures 5-10 show a 3D plot (obtained from MLPNN) of the effects of two different independent
nu
variables on pressure drop while the other factors kept constant. As shown in these figures each
variable can affect pressure drop whereas as mentioned in Figure 1, the effectiveness of these
Ma
variables can vary. As expected, it obviously shows an increase in the consistency coefficient,
liquid velocity, and flow behavior index, the pressure drop significantly increases. However,
ed
increasing the other variables do not have a significant effect on the pressure drop.
pt
ce
Ac
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
Figure 6. Three-dimension plots of pressure drop based on the pipe diameter and gas superficial
Figure 8. Three-dimension plots of pressure drop based on the flow behavior index and
Figure 10. Three-dimension plots of pressure drop based on the liquid density and liquid
This section is dedicated to the evaluation of the ability of the proposed MLPNN with the optimum
topology from alternative methods. For this purpose, two common regression models, including
d
te
The results of MLPNN (training + test) for pressure drop are compared with four statistical indices,
namely AARD%, MSE, RMSE, and R2 with other models in Tables 4. This Table proves the higher
di
accuracy of MLPNN, compared to the other models, in the prediction of pressure drop.
e
py
Table 4 Comparison of the accuracies of MPLNN and other models.
Co
Statistical index
Model dataset
AARD% MSE RMSE R2
Train 4.36 0.0018 0.0424 0.99596
ot
MLPNN Test 5.83 0.0064 0.0803 0.98622
Total 4.58 0.0025 0.0500 0.99438
tN
Train 4.69 0.0022 0.0474 0.99486
CFNN Test 5.75 0.0039 0.0622 0.99310
Total 4.85 0.0025 0.0499 0.99449
rip
Moreover, the mean squared errors (MAE) obtained from the three proposed empirical models in
ce
this study were compared with MLPNN in Tables 5. It shows that the most accurate results belong
Ac
to the MLPNN that comply well with the experimental data. The results of these studies revealed
that the behavior of gas/non-Newtonian liquid flow in pipes was different which affects the fluids
transportation and pertinent industrial operations.
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
d
te
4. Conclusions
di
The ANN modeling was applied to predict the pressure drop in gas/non-Newtonian liquid to
e
investigate the impacts of physical and operating conditions on two-phase flow. The design of the
py
MLPNN with a 6-10-1 structure was carried out using 511 data that were experimentally obtained.
Co
The training was done by 85% of these data that were selected randomly for network training, the
validation process was performed by 15% of the remaining data. The actual performance function
ot
of the resulted neural network model was inspected by the rest of the data points. The estimation
of pressure drop using the optimum MLPNN has resulted in the values of 4.58%, 0.0025, 0.05,
tN
and 0.99438 for AARD, MSE, RMSE, and R2, respectively. These statistical values and error
rip
indices prove the reliability of the MLPNN for the prediction of pressure drop in pipes among the
other empirical correlations and various AI-based approaches.
sc
Nomenclature
nu
b Bias
Ma
N Number of datasets
u velocity [m s–1]
w Weight
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
Abbreviations
AI Artificial intelligence
d
te
GRNN Generalized regression neural network
di
MSE Mean squared errors
e
MAE Mean absolute error
py
RBFNN Radial basis function neural networks
Co
R2 Regression coefficient
ot
Subscripts/superscripts
tN
g Gas phase
gs Gas slug
rip
l Liquid phase
ls Liquid slug
sc
m Mixture phases
nu
tp Two-phase
max Maximum
ed
min Minimum
Greek letters
ce
k Ratio of the liquid slug zone length to the slug unit length [–]
Acknowledgment
The first author acknowledges the support provided by Alexander von Humboldt Foundation
through the project FRA-1204799-HFST-E for the experienced researcher. The authors also
d
te
References
di
[1] Moayedi H, Aghel B, Vaferi B, Foong LK, Bui DT. The feasibility of Levenberg–
e
Marquardt algorithm combined with imperialist competitive computational method
py
predicting drag reduction in crude oil pipelines. J Pet Sci Eng 2020;185:106634.
Co
[2] Sorgun M, Murat Ozbayoglu A, Evren Ozbayoglu M. Support vector regression and
computational fluid dynamics modeling of Newtonian and Non-Newtonian fluids in
ot
annulus with pipe rotation. J Energy Resour Technol 2015;137.
tN
[3] Ferrari M, Bonzanini A, Poesio P. A slug capturing method in unconventional scenarios:
The 5ESCARGOTS code applied to non-Newtonian fluids, high viscous oils and complex
rip
[4] Almani S, Haydar A, Blel W, Gadoin E, Gentric C. Thin gap bubble column with a non-
Newtonian liquid phase: study of the hydrodynamics and gas-liquid mass transfer, 2019.
nu
[5] Khatib Z, Richardson JF. Vertical co-current flow of air and shear thinning suspensions of
Ma
[6] Dziubinski M, Fidos H, Sosno M. The flow pattern map of a two-phase non-Newtonian
ed
shear-thinning fluid flows in inclined smooth pipes. Int J Multiph Flow 2007;33:948–69.
[8] Jumpholkul C, Asirvatham LG, Dalkılıç AS, Mahian O, Ahn HS, Jerng D-W, et al.
Ac
Experimental investigation of the heat transfer and pressure drop characteristics of SiO
2/water nanofluids flowing through a circular tube equipped with free rotating swirl
generators. Heat Mass Transf 2019:1–14.
and pressure drop of R-134a flowing in a dimpled tube. Int J Heat Mass Transf
2017;106:437–48.
d
[11] Almasi F, Shadloo MS, Hadjadj A, Ozbulut M, Tofighi N, Yildiz M. Numerical
te
simulations of multi-phase electro-hydrodynamics flows using a simple incompressible
di
smoothed particle hydrodynamics method. Comput Math with Appl 2019.
e
py
[12] Bhagwat SM, Ghajar AJ. Experimental investigation of non-boiling gas-liquid two phase
flow in upward inclined pipes. Exp Therm Fluid Sci 2016;79:301–18.
Co
[13] Liu Z, Liao R, Luo W, Ribeiro JXF, Su Y. Friction pressure drop model of gas-liquid two-
phase flow in an inclined pipe with high gas and liquid velocities. AIP Adv 2019;9:85025.
ot
[14] Rahmat A, Tofighi N, Yildiz M. The combined effect of electric forces and confinement
tN
ratio on the bubble rising. Int J Heat Fluid Flow 2017;65:352–62.
rip
[16] Nwaka N, Wei C, Chen Y. A Simplified Two-Phase Flow Model for Riser Gas
nu
[18] Cheng L, Ribatski G, Thome JR. Two-phase flow patterns and flow-pattern maps:
pt
[19] Shannak BA. Frictional pressure drop of gas liquid two-phase flow in pipes. Nucl Eng Des
Ac
2008;238:3277–84.
[20] Garoosi F, Bagheri G, Rashidi MM. Two phase simulation of natural convection and
mixed convection of the nanofluid in a square cavity. Powder Technol 2015;275:239–56.
[21] Garoosi F, Rashidi MM. Two phase flow simulation of conjugate natural convection of
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
the nanofluid in a partitioned heat exchanger containing several conducting obstacles. Int J
Mech Sci 2017;130:282–306.
[22] Chhabra RP, Richardson JF. Prediction of flow pattern for the co-current flow of gas and
[23] Mandhane JM, Gregory GA, Aziz K. A flow pattern map for gas—liquid flow in
d
horizontal pipes. Int J Multiph Flow 1974;1:537–53.
te
[24] Dziubinski M. A General Correlation for 2-Phase Pressure-Drop in Intermittent Flow of
di
Gas and Non-Newtonian Liquid-Mixtures in a Pipe. Chem Eng Res Des 1995;73:528–34.
e
py
[25] Ruiz-Viera MJ, Delgado MA, Franco JM, Sánchez MC, Gallegos C. On the drag
reduction for the two-phase horizontal pipe flow of highly viscous non-Newtonian
Co
liquid/air mixtures: Case of lubricating grease. Int J Multiph Flow 2006;32:232–47.
[26] Heywood NI, Charles ME. The stratified flow of gas and non-Newtonian liquid in
ot
horizontal pipes. Int J Multiph Flow 1979;5:341–52.
tN
[27] Taitel Y, Dukler AE. A model for predicting flow regime transitions in horizontal and
rip
[28] Eisenberg FG, Weinberger CB. Annular two-phase flow of gases and non-Newtonian
sc
[29] Alizadehdakhel A, Rahimi M, Sanjari J, Alsairafi AA. CFD and artificial neural network
Ma
modeling of two-phase flow pressure drop. Int Commun Heat Mass Transf 2009;36:850–
6.
ed
[30] Osman E-SA. Artificial neural network models for identifying flow regimes and
predicting liquid holdup in horizontal multiphase flow. SPE Prod Facil 2004;19:33–40.
pt
[31] Xie T, Ghiaasiaan SM, Karrila S. Artificial neural network approach for flow regime
ce
[32] Cai S, Toral H. Flow rate measurement in air-water horizontal pipeline by neural
networks. Proc. 1993 Int. Conf. Neural Networks (IJCNN-93-Nagoya, Japan), vol. 2,
IEEE; 1993, p. 2013–6.
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
[33] Osman E-SA, Aggour MA. Artificial neural network model for accurate prediction of
pressure drop in horizontal and near-horizontal-multiphase flow. Pet Sci Technol
2002;20:1–15.
d
2002.
te
[35] Chhabra RP, Richardson JF. Non-newtonian fluid behaviour. Non-newtonian flow and
di
applied rheology 2008.
e
py
[36] Taitel Y, Barnea D. A consistent approach for calculating pressure drop in inclined slug
flow. Chem Eng Sci 1990;45:1199–206.
Co
[37] Xu J, Wu Y, Li H, Guo J, Chang Y. Study of drag reduction by gas injection for power-
law fluid flow in horizontal stratified and slug flow regimes. Chem Eng J 2009;147:235–
44.
ot
tN
[38] Xu J. A simple correlation for prediction of the liquid slug holdup in gas/non-Newtonian
rip
fluids: Horizontal to upward inclined flow. Exp Therm Fluid Sci 2013;44:893–6.
[39] Komeilibirjandi A, Raffiee AH, Maleki A, Nazari MA, Shadloo MS. Thermal
sc
[40] Zheng Y, Shadloo MS, Nasiri H, Maleki A, Karimipour A, Tlili I. Prediction of Viscosity
Ma
of Biodiesel Blends Using Various Artificial Model and Comparison with Empirical
Correlations. Renew Energy 2020.
ed
[41] Maleki A, Elahi M, Assad MEH, Nazari MA, Shadloo MS, Nabipour N. Thermal
pt
[42] Aghel B, Rezaei A, Mohadesi M. Modeling and prediction of water quality parameters
using a hybrid particle swarm optimization–neural fuzzy approach. Int J Environ Sci
Technol 2018:1–10.
Technol 2016;138.
[44] Moayedi H, Aghel B, Foong LK, Bui DT. Feature validity during machine learning
paradigms for predicting biodiesel purity. Fuel 2019:116498.
d
Therm Sci Eng Prog 2018;6:226–35.
te
[46] Wadkar D, Kote A. Prediction of residual chlorine in a Water treatment plant using
di
Generalized regression neural Network. Int J Civ Eng Technol 2017;8:1264–70.
e
py
[47] Ghosh T, Martinsen K, Dan PK. Data-Driven Beetle Antennae Search Algorithm for
Electrical Power Modeling of a Combined Cycle Power Plant. World Congr. Glob.
Co
Optim., Springer; 2019, p. 906–15.
[48] Farooqi SI, Richardson JF. Horizontal flow of air and liquid (Newtonian and non-
ot
Newtonian) in a smooth pipe. Part II: Average pressure drop. Trans IChemE
tN
1982;60:323–33.
rip
[49] Chhabra RP, Richardson JF, Farooqi SI, Wardle AP. Co-current flow of air and shear
thinning suspensions in pipes of large diameter. Chem Eng Res Des 1983;61:56–61.
sc
[50] Wu B. CFD simulation of gas and non-Newtonian fluid two-phase flow in anaerobic
nu
[51] Mowla D, Naderi A. Experimental study of drag reduction by a polymeric additive in slug
two-phase flow of crude oil and air in horizontal pipes. Chem Eng Sci 2006;61:1549–54.
ed
[52] Li H, Wong TN, Skote M, Duan F. Non-Newtonian two-phase stratified flow with curved
interface through horizontal and inclined pipes. Int J Heat Mass Transf 2014;74:113–20.
pt
[53] Firouzi M, Hashemabadi SH. Exact solution of two phase stratified flow through the pipes
ce
2009;36:768–75.
[54] Vaferi B, Samimi F, Pakgohar E, Mowla D. Artificial neural network approach for
prediction of thermal behavior of nanofluids flowing through circular tubes. Powder
Technol 2014;267:1–10.
Journal of Energy Resources Technology. Received April 27, 2020;
Accepted manuscript posted May 20, 2020. doi:10.1115/1.4047593
Copyright © 2020 by ASME
[55] Demuth HB, Beale MH, De Jess O, Hagan MT. Neural network design. Martin Hagan;
2014.
[56] Xu J, Gao M, Zhang J. Pressure Drop Models for Gas/Non-Newtonian Power-Law Fluids
d
te
e di
py
Co
ot
tN
rip
sc
nu
Ma
ed
pt
ce
Ac