You are on page 1of 30

Journal Pre-proof

Prediction of multiphase flow rate for artificially flowing wells using rigorous artificial
neural network technique

Reda Abdel Azim

PII: S0955-5986(20)30176-X
DOI: https://doi.org/10.1016/j.flowmeasinst.2020.101835
Reference: JFMI 101835

To appear in: Flow Measurement and Instrumentation

Received Date: 14 June 2020


Revised Date: 26 August 2020
Accepted Date: 16 September 2020

Please cite this article as: R.A. Azim, Prediction of multiphase flow rate for artificially flowing wells using
rigorous artificial neural network technique, Flow Measurement and Instrumentation (2020), doi: https://
doi.org/10.1016/j.flowmeasinst.2020.101835.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.

© 2020 Published by Elsevier Ltd.


Dear Editor and anonymous reviewers.

I would like to thank the reviewers for their thoughtful comments and efforts towards improving
the manuscript. In the following, I addressed comments specific to each reviewer below. In
addition, the new sections added in this manuscript are highlighted with red colour.

Regards

Reda Abdel Azim


Assistant Professor, Petroleum Engineering Department
College of Engineering

of
ro
-p
Zakho Road, 42001 Duhok, Kurdistan Region of Iraq
Reda.abdulrasoul@auk.edu.krd
re
Facebook | Twitter | Instagram | YouTube | LinkedIn
lP
na
ur
Jo
Prediction of multiphase flow rate for artificially
flowing wells using rigorous artificial neural
network technique
Reda Abdel Azim

American University of Kurdistan, Petroleum Engineering Department, Duhok, Iraq

Abstract

of
This main aim of this study is to generate an intensive artificial neural network model (ANN) based

ro
on FORTRAN language to develop a physical equation for oil rate prediction in wells lifted by ESP
-p
pumps. The backpropagation algorithm (BP) is selected in this study as a learning algorithm with
re
its sigmoid curve based on the comparison performed against scaled conjugate gradient (SCG) and
lP

one-step secant (OSS) algorithms.

300 data points are collected from 2 fields in Gulf of Suez Egypt used in the ANN model. The results
na

show that the optimum distribution for the collected data is of 70% and 30% for training and
ur

testing processes, respectively. This distribution yields the highest R2 of 0.988 and lowest mean

square error of 0.025. Furthermore, based on the statistical analysis presented in this study, it has
Jo

been found that the optimum number of hidden layers and neuron are one layer and two neuros,

respectively.

The newly ANN and correlation can predict the oil rate at the surface with accuracy exceeding 96%

and that is extremely efficient. A comparison is conducted between the presented correlation in this

study and other published correlations (Gilbert and Ros correlations) based on R2 value and mean

square error. The results show that the new correlation has the highest R2 value with the lowest

mean square error.

Keywords: Artificial neural network; oil flow rate; Correlation; wellhead and pump parameters

1
1. Introduction

The multiphase flow occurs occasionally in oil wells during the life span of the oil field [1]. The most

common problems associated with multiphase flow regime is the precise determination of the

liquid flow rate, wellhead pressure and choke size. According to determination of choke size and

well head pressure, the issue is easy determined with using well-calibrated pressure gauge.

However, determining the accurate surface production rate has a significant importance when it

of
comes to field development, dealing with gas/water conning and sand production. Thus, this

ro
achieved by controlling the surface flow rate through using of choke that supply sufficient back

pressure to adjust production rates [2]. -p


re
Likewise, the choke at the wellhead perform critical functions include: (a) preventing slug flow and

consequently protect wellhead assembly; (b) restraint fluctuations occur in the downstream that
lP

affect upstream pressures. When the flow velocity through choke near the speed of sound, the
na

downstream pressure disturbance does not have a significant effect on upstream pressures, and

therefore, the choke will maintain the flow in this region.


ur
Jo

The surface well production rate determination has utmost importance in specifying and

monitoring wells performance. Test separators and flow meters are employed in oil field to

measure individual well contribution to the field production. Test separators provide accurate

measurements; nevertheless, they are very expensive and requires regular maintenance [3]. In

addition, using of flow meters during the well test operations need to be calibrated using test

separator data that led to inaccurate flow rates and consequent erroneous calculations [4]. There

are plenty of flow correlations can be found in the literature for rate estimation at the wellhead. The

correlations provided can be divided into two categories: theoretical, which require several inputs

and hence are usually impractical to use, and empirical, which are generated for a limited range of

2
data. In addition, not more correlation can be found for artificially flowing wells specially for ESP

pump wells.

Tangren [5] presents the first multiphase flow theory for the choke. The theory used when the

produced phase is continuous and the ratio of gas to liquid is less than one. Tangren [5] theory has

been modified by Gilbert [6] for naturally flowing wells followed by other developments as

Baxandall [7] and Ros [8].

Gilbert [6] presented a correlation (see Eq.1) for critical flow rate calculation using 268 data points

of
include choke size range from 6/18 to 64/64 inches.

ro
= (1)
-p
Where
re
Pwh is the wellhead pressure (psia),
lP

D64 is the size of the choke (1/64 inch),


na

GLR is the ratio of gas to liquid (Standard cubic feet/Stock tank barrel), and,

QL is the critical-flow liquid rate (Stock tank barrel per day).


ur

A, B and C are the main coefficients evaluated based on sufficient data is available for certain
Jo

reservoir. A= 0.1, B= 1.89 and C= 0.546.

Baxandall modified the Gilbert’s empirical correlation for the critical rate calculation and Ros [8]

presents a new correlation to derive oil and gas mass under critical flow conditions. Achong [9]

proposed a new set of values for coefficients A =0.2618, B =1.88 and C=0.65 of Gilbert’s equation.

Poettman [10] improved the Ros’s correlation by using collected data from an oil field with choke

restrictions.

Other studies present their correlation based on Gilbert’s equation as Osman and Dokla [11]

proposed a correlation to determine the flow rates through wellhead using least square method for

gas condensate reservoirs. Two phase flow equation for Perkins [12] has been developed based on

3
momentum, mass, and energy balance equation to demonstrate isentropic flow through

restrictions.

The presented literature review indicates that numerous multiphase flow correlations have been

developed to predict the wellhead flow rate through choke. Those correlations ignored significant

well and pump parameters especially for artificially flowing wells using ESP pumps.

The purpose of this study is to present a rigorous correlation using artificial neural network

technique combined with back propagation learning algorithm to predict the oil rate for artificially

of
flowing wells. This correlation includes numerous newly incorporated parameters such as pump

ro
intake/discharge temperature, choke size, upstream and downstream pressures through choke,

-p
pump intake and discharge pressures, casing head pressure, gas oil ratio and water cut.
re
300 data points are used in this study to achieve the goal. The data are collected from 7 oil wells in

Gulf of Suez fields, Egypt. In addition, to avoid the deficiencies exist in previous presented ANN
lP

models, a new convergence scheme presented in this study.


na

2. Artificial neural network


ur

The artificial neural network ANN models are used in several petroleum engineering applications
Jo

include diagnosis of the drill bit operation [13], well test model analysis and interpretation[14],

ESP and PCP pumps card diagnosis [15] and identification of well performance [16]. The neural

network calculation using two passes, forward pass used for the forecast purposes and backward

pass is for the training targets. The structure of the neural network consists of input, hidden,

output layers, neurons, objective, transformation functions and training algorithm. In training

process, a comparison is performed between the target values the actual data. The difference

between calculated and target outputs is assessed to characterize the ANN error (also known as

loss value). The error that generated for every training record are used in calculating the objective

function. Then, the training learning algorithm is used to reallocate the total ANN errors onto the

4
weight that connect all nodes (e.g. back propagation learning algorithm). The weight adjustment

process is controlled by learning and decay rates. The learning rate is used in controlling the

extreme weight changes in the training process while the decay rate is used in controlling the

lowest change in the weight value. The optimization algorithm is utilized as well in the neural

network structure to minimize the error of several training records until ANN weights attain error

tolerance. A bias node in the hidden layer is used to scale up the weighted sum to its value.

2.1 Learning algorithm

of
The back-propagation learning algorithm is widely used in the learning process for the supervised

ro
neural net types. These types of algorithms require pre-existing training patterns and involve

-p
forward and backward steps. The forward step sends signals of input data via the neurons of each
re
layer to calculate the actual targets. Then, the step of backward propagation used to calculate the

error path using error value between actual and target outputs. New sets of weights are iteratively
lP

computed for each layer backwards using the targets (output), by adjusting the existing weights
na

based on the calculated error values [17]. The root-mean-square error (MSE) is normally used as a

measure of the global error [18], which can be defined as:


ur
Jo

∑ ∑ (x p − yp )
n1 n2 2

MSE = 1 1
(2)
n1.n2

Where n1 is the number of training patterns, n2 is the number of output neurons in the output layer.

Values of xp and yp are the target and calculated output values, respectively. The neural network

structure includes the weights that are automatically adjusted slowly to allow the network to

generalize among the input patterns in the training set.

5
2.2 Forward Propagation

The forward propagation starts when input signals received by the neurons in the input layer. Input

neurons do not operate any mathematical task. The mathematical operation begins within the

hidden layer neurons. Each neuron in the hidden layer calculates the linear combination of the

input and weight vectors. Fig.1 shows the details of this step.

n
S = ∑ xi wi

of
i=0 (3)

ro
Where n is the dimensions of the input vector X (x0, x1,..., xn), and W (w0,w1,...,wn) is corresponding

-p
weight vector. Note that w0 is the weight for the bias input x0 (which is equal to one). The summed
re
value S is then passed through a transfer function. The types of functions that can be used include
lP

logistic, hard limit, linear and linear threshold. The most commonly employed function in

backpropagation is a sigmoid curve called the logistic function. This function squashes the summed
na

inputs to fit within certain limits. The forward propagation step continues until the input signals reach
ur

the output layer. The most employed function in backpropagation algorithm is a sigmoid curve

called the logistic function with the form:


Jo

1
f (s) =
1+ e−s (4)

The forward propagation step continues until the input signals reach the output layer.

6
Xo Wo

n
S = ∑ xi wi F (S)
i=0

Wn

Xn

of
Fig.1 Mathematical operations of hidden neuron

ro
-p
re
2.3 Backward Propagation
lP

The back-propagation step used the output values that are obtained at the output layer and

compare them with actual targets. If the difference is calculated, then the network sends the error
na

backwards through the hidden and input layers. Therefore, this algorithm called backpropagation
ur

(BP) and it used as learning algorithm for the neural network. The error values are then utilized to
Jo

update the weights in such a way that the errors will be reduced in the next iteration.

1 n2
∑ ( xp − y p )
2
Ep = (5)
2 1

Where E is the error vector for the training pattern p

To modify the weights on the connections, a learning rule, called the delta rule or gradient descent

method [18], is used to compute a "delta weight" vector.

7
2.4 Network training

2.4.1 Learning Speed

“The convergence speed of the network is improved significantly by an acceleration method in

which the speed of the point is modified in the weight space instead of its position This keeps the

weight changes on a faster and more even path by adding a fraction of the previous weight changes

“[19]:

(6)
w ( t + 1) = w(t ) + β [ ∆w(t ) ] + α [ w(t − 1) ]

of
ro
“Where α is the energy constant, w is the weight, ∆ is the weight difference, t in incremented by 1

for each epoch, and -p


is the learning constant. This constant is used to effectively increase the step
re
size to reduce abrupt gradient changes. The learning and momentum constants are set in a range of

0 and 1” [19].
lP

2.4.2 Network performance


na

To check the performance of the neural net, the MSE presented in Eq. 2 is used to indicate the

performance. Thus, a so-called learning curve, which it shows the MSE versus the number of
ur

epochs. If the performance is lower than an appropriate level, this means that the network is not
Jo

well trained, and one may re-train the network by either adjusting some of the network parameters

(e.g. the number of hidden neurons), or by changing the variations present in the training patterns.

2.5 Application of ANN in predicting the oil flow rate

Mirzaei-Paiaman and Salavati [20] applied artificial neural network technique for oil flow rate

prediction through choke by using the upstream pressure and size of choke, and gas to oil ratio.

Gorjaei et al. [21] and Nejatian et al [22] used Least Square Support Vector Machine (LSSVM) for

modeling critical flow through choke. Ghareeb et al. [23] generated a new correlation for artificially

flowing wells to predict surface flow rate based on measured wellhead pressure, temperature, GLR

8
and water cut using the genetic algorithm technique. Khorzoughi et al. [24] proposed a multiphase

flow chock correlation using non-linear optimization method and Monto Carlo simulation. Seidi

and Sayahi [25] use a combination of genetic algorithm and non-linear regression to forecast the

sub-critica flow through chokes.

Leal [26] presented an analytical approach to estimate the flow rate for Al-Ghwar field gas

condensate wells. Ghorbani [27] presented an optimization strategy using genetic algorithm based

nonlinear optimization to determine liquid flow rate. The results show that predicted flow rates

of
were very close to the actual wells rate with R2= 0.997.

ro
Al Ajmi et al. [28] used a fuzzy logic to develop a robust model to estimate oil flow rate as a function

-p
of certain operational conditions and the results reported absolute error of 13.92%, the work
re
concentrating on sonic and subsonic conditions. Hasanvand and Berneti [29] used a neural network

technique to predict oil flow rate for a data set collected from a North Iranian oil field. They
lP

concluded that neural net performed well without using uncertain data from multiphase flow
na

metering.

Naseri et al. [30] used a combination of optimization technique and genetic algorithm for accurate
ur

estimation of orifice choke coefficients for gas flow. Choubineh et al. [31] used a neural network
Jo

model in combination with an optimization algorithm for gas flow rate prediction. The model based

on 113 inputs data set collected from 12 Iranian oil fields. The results show absolute average error

between actual and predicated rates of 2%.

Going through the previous work presented in the literature review, it has been observed the

carried work did not provide a single physical equation that can be used to estimate oil flow rate for

artificially flowing wells. Therefore, the main aim of this work is to provide a novel empirical

correlation based on ANN technique by clarifying the black-box concept.

Additionally, in this study, two parameters are used to speed up the convergence between actual

and predicted data: step size and momentum [32]. The speed of convergence in this study is

9
improved by using an acceleration method. This will assist in getting fast weight changes by adding

fraction to previous weight changes as explained in section 2.4.1.

3. Data Analysis

3.1 Data description

The data sets of 300 points are collected from Gulf of Suez fields in Egypt [33]. The data contains 10

inputs used for the training process include; choke upstream, downstream pressures (psi), ESP

of
pump intake , discharge pressures (psi), intake and discharge pump temperature (F), wellhead

ro
pressure (psi), GOR (scf/stb), choke size (in) and water cut. The target output is the oil flow rate
-p
(bbl/D). The input and output data are normalized a range of 0 and 1 as follows:
re
∗ ∗
= ∗ ∗ (7)
lP

“Where y is the normalized data, y* is the raw data, y*min and y*max are the minimum and
na

maximum values of the raw variable, respectively” [19].


ur

The input dataset is filtered before beginning in the training process by removing the unrecorded
Jo

values of choke size and daily flow rate. Table 1 shows the statistical analysis of the input data.

Table 1 Statistics of the input data

Parameter Choke GOR Upstream Downstre Pump Pump Pump Pump Water Wellhead
pressure am intake dischar intake discharge cut pressure
size (in) (scf/stb) (psi) pressure pressur ge temperat temperat
(psi) e (psi) pressu ure (F) ure (F)
re (psi)
Min 12 210 100 10 175 322 92 99 0 0

Max 64 360 650 220 883 2016 180 161 0.043 600

Standard 12.4 115.66 95 29 241.087 392.81 5.8955 14.755 0.015 24.960

10
deviation

Skewnes 0.00583 0.021 0.28431 -1.28 1.3455 0.1549 1.607 0.309 7.38 1.439

Mean 64 290 350 190 815 814 122 158 0 200

3.2 Statistical analysis

3.2.1 Selection of the best distribution of input data and learning algorithm

of
To choose the optimum distribution of the gathered data for functioning training and testing

operations, a statistical analysis is performed using different percentages of data distributed

ro
between training and testing. The results show that 70% for training and 30% for testing gives the
-p
highest R2 value and lowest mean square error, as shown in Table 2.
re
lP

Table 2 Analysis of training and testing data distribution to be used in ANN model

Trial Distribution R2Training MSE R2 testing MSE


na

Training Testing
ur

1 10 90 0.62 0.58 0.57 0.62


Jo

2 20 80 0.742 0.43 0.752 0.41

3 30 70 0.821 0.354 0.839 0.378

6 60 40 0.885 0.263 0.861 0.24

7 70 30 0.988 0.026 0.956 0.0201

8 80 20 0.882 0.095 0.878 0.097

9 90 10 0.87 1.03 0.845 1.102

Three learning algorithms are tested in this study to select the optimum one for training and testing

processes. These algorithms are: (1) Back propagation (BP), (2) scaled conjugate gradient (SCG),

and (3) one-step secant (OSS). The results show that the BP algorithm is the best learning one

based on R2 of 0.988 and MSE = 0.025 comparing to other algorithms as shown if Fig.2. The SCG

11
algorithm yields R2 of 0.84 with MSE of 0.124 while OSS algorithm resulted in R2 of 0.815 with MSE

of 0.287 (see Fig.2).

of
ro
-p
re
Fig.2. The results of R2 (a) and MSE (b) for BP, SCG, and OSS functions.
lP

Several ANN model scenarios are performed to study the effect of each input parameter on the flow
na

rate calculation. In each trial, the influence of each input parameter is detected while other

parameters are kept constant. Table 3, three of these runs is presented as an example. the
ur

correlation coefficient and the mean square error were used to assess the accuracy of each ANN
Jo

run.

Table 3. Results of selected trials performed to study the effect of input parameters on Qo.

Trial Input parameters R MSE R MSE

number training testing

8 Pups,PDs,Pint,Pdisc,Tint,Tm,WHP,ChokeS,WC,GOR 0.99 0.015 0.98 0.0175

12 Pups,PDs,Pint,Pdisc,Tint,Tm,WHP,ChokeS,WC 0.954 0.154 0.932 0.168

13 Pups,PDs,Pint,Pdisc,Tint,Tm,WHP,WC,GOR 0.946 0.245 0.9302 0.278

12
3.2.2 Optimizing the number of hidden layers and neurons

To optimize the number of layers and neurons used in the hidden layer, an optimization algorithm

used here by performing iterative process based on mean square error (MSE) (see Eq.2) and

Levenberg-Marquardt algorithm [34]. The results show that the optimum number of hidden layers

is one layer while optimum number of neurons is 15 neurons based on the MSE = 0.02 (see Fig.3)

of
ro
-p
re
lP
na

Fig.3. (a) MSE vs. Number of Neurons in the Hidden Layer and (b) MSE vs. Number of Hidden
ur

Layers
Jo

4. Results and discussion

In total of 300 points for wells flowing artificially using ESP pumps are used in the neural net, in

which 70% of the data is used for training and 30% used for testing. Results of the ANN model

shows a reasonable match between the actual and predicted oil flow rate as shown in Figs. 4 and 5.

The R2 value for the training process is 0.988 (see Fig.4) while the testing process gives R2 of 0.95

(see Fig.5). Fig.6 also show that an excellent match between the predicted and actual oil flow rate as

a function of the measured surface casing head pressure. This exceptional match is obtained based

on the BP learning algorithms used in the developed neural network for this study. Table 4 shows

the data used for testing operation. In addition, Fig.7 shows the performance of ANN model in this

13
study. As shown from the figure that only 500 epochs (iterations) required to get the minimum

error which confirm the effectiveness of the convergence technique used in this study.

1.0

0.8

of
0.6

ro
0.4

0.2 -p
re
0.0
0.0 0.5 1.0
lP

Normalized predicted oil rate (bbl/D)


na

Fig.4. Actual Qo versus targets using training data


ur
Jo

Fig.5. Actual Qo versus targets using testing data

14
Table 4. Data set used for the testing operation.

UP/D pres., psi Pump Pi, Pump Pump Ti, Pump CHP, psi Choke WC
psi Pd, psi ºF Tm, ºF
390 200 315 1243 133 151 200 26 0
390 200 304 1239 133 151 200 24 0
270 220 200 1149 132 151 210 32 0
340 220 196 1188 132 151 220 32 0
330 200 199 1187 132 151 210 32 0
320 210 201 1170 132 150 210 32 0
300 200 201 1169 132 151 210 32 0
325 200 195 1198 132 150 190 32 0
310 200 161 1182 130 149 130 32 0
340 200 158 1212 130 149 130 26 0

of
320 200 155 1200 130 129 130 26 0.078
330 200 153 1210 130 149 130 26 0.073

ro
330 200 129 1198 130 148 70 32 0.073
320 190 127 1195 130 148 70 32 0.073
310
320
200
190
126
108
1195
1193
130
130
-p 148
148
70
20
32
32
0.073
0.073
re
310 190 104 1190 129 147 20 32 0.11
300 190 93 1189 129 147 20 32 0.11
300 180 89 1181 129 147 20 32 0.13
lP

295 180 86 1178 129 147 20 32 0.13


310 170 47.6 1210 159 147 20 32 0.13
310 170 48 1207 130 148 20 32 0.17
na

310 170 47 1209 130 147 20 32 0.16


330 170 48 1239 130 148 20 32 0.16
330 170 48 1230 130 148 20 32 0.19
ur

320 170 48 1224 130 148 20 32 0.19


320 170 48 1205 130 148 20 32 0.2
Jo

300 170 49 1005 137 155 20 32 0.2


310 170 47 1220 130 148 20 32 0.2
320 170 48 1224 130 148 50 32 0.2
310 170 48 1223 130 148 20 32 0.2
320 165 49 1242 132 146 20 32 0.2
310 160 49 1236 131 148 20 32 0.2
320 165 49 1230 131 148 20 32 0.2
300 130 47 1213 131 149 20 32 0.2

15
of
ro
-p
Fig.6. Actual Qo versus predicted values for testing data.
re
lP
na
ur
Jo

Fig.7. Performance of ANN model Vs epochs

16
As observing the encourage outcomes from the training and testing operations, the work in this

study persisted to develop a correlation that can be used without the need of ANN software. The

proposed equations achieved based on set of weights and biases relate to input and output layers.

The weights and biases are given in Table 5.

of
ro
-p
re
lP
na
ur
Jo

17
Table 5. ANN model weights and biases Eq. (8).

Hidden Connecting Weights (w1) Connecting Hidden Output

layer Weights layer layer

neuron (w2) bias bias

o f
(b1) (b2)

ro
1 3.316916

-p
-0.63041 -0.83904 -0.65756 -0.56723 0.47278 -7.02E-02 7.39E-02 -0.4797 -1.18736 -0.34826 4.74E-02 -0.71169

re
2
-1.33095 -0.51257 0.114049 -0.46938 -0.84618 -0.24849 -8.98E-03 -0.64943 0.131325 -0.12373 -0.25674 -1.75914

lP
3
2.007709 -0.70823 -4.7138 -0.37651 -3.58E-02 -1.19279 4.64E-02 -0.18741 -2.22794 7.67E-03 0.290055 -0.81135

na
4
-0.50668 -0.16721 0.128484 0.164382 0.895096 -0.54222 -0.11059 -0.2792 8.06E-02 0.188385 -3.62141 -1.35397

5 -1.07E-
ur
Jo
-0.39775 02 0.755416 -0.43032 -1.69232 -0.47545 -0.14784 -0.27899 0.10835 -0.99319 0.54645 -1.69456

6
-0.49229 0.220615 -1.57242 -0.91807 -8.31E-02 -0.61624 -0.18823 0.136869 4.59E-02 3.094312 -6.57418 -1.37307

7
-0.69166 0.134708 -0.4291 0.239182 -0.52313 -0.68037 -0.59963 -0.76975 -6.65E-02 0.266372 5.94E-02 0.558922

8
-0.40112 -1.83155 -0.32786 -0.53752 -2.11205 -0.584 0.132848 -0.57058 -1.48E-02 -0.30618 -0.11738 -1.28718

9
1.091616 3.074556 -4.41E-02 -0.2078 -3.76874 -0.29782 0.583843 -0.63247 -0.11391 1.765111 3.64E-03 -0.92801

10
-0.26715 -0.11796 0.630029 -4.19467 -0.36922 -0.30395 -2.27356 -0.47881 0.230735 -0.74395 0.15775 -1.07131

11
-7.42E-02 6.62E-02 -2.86E-02 0.973263 0.357874 -7.15E-02 -0.94815 -0.83404 -0.31806 -0.28159 -0.20597 2.831101

18
12
0.475648 0.127874 -0.81123 0.125377 0.30065 -0.14107 -0.96619 -0.1953 0.179283 1.111362 -5.52E-02 -1.16456

13
-0.23757 -0.12901 0.558702 0.166034 -0.46222 -0.51529 6.44E-02 -0.12247 -1.05514 0.356071 0.743097 -2.32931

14
-0.60569 -0.14874 -0.35646 6.30E-02 -0.90651 3.15E-02 -0.4314 -0.31132 -1.3149 -0.84608 -0.27947 -1.44735

15 -9.24E-

-1.09884 -2.40777 -0.24774 1.45E-02 -1.21522 1.827596 -0.52242 02 -3.01712 0.788413 3.580081 -2.82884

o f
ro
The novel correlation developed by using ANN technique for oil flow rate estimation in artificial lift oil wells using ESP pumps is:

-p
  
N  

re
1
Qon =  ∑ w2,i  −1 +b 2

i =1

(( ) ) 
− Pups×w1i ,1+ PDs×w1i ,2 + P in×w1i ,3 + Pdisc×w1i ,4 +T int ×w1i ,5 +Tm×w1i ,6 +WHP×w1i ,7 +ChokeS ×w1i ,8 +WC×w1i ,9 +GOR×w1i ,10 +b1  (8)

lP

  1+exp 

na
ur
Jo

19
Where Qon is the normalized oil flow rate, (w1, i) is the weight vector linking the input and the

hidden layer, i is the index of the neuron in the hidden layer, b1 is the biases vector for the input

layer, and b2 for the output layer, Upstream pressure (Pups), Downstream pressure (PDs), ESP

intake (Pint), ESP discharge (Pdisc), ESP intake temperature (Tint), ESP discharge temperature

(Tm),Well head pressure (WHP), Choke size (ChokeS), Water cut (WC) and Gas to Oil ratio (GOR)

Finally, the required Qo can be obtained by deformalizing Qon as follows:

of
Qo = 6172.1*Qon + 266.1 (9)

ro
4.1 Validation of the Proposed Model
-p
re
To confirm the validity of the new developed correlation, a comparison is performed against well-
lP

known published correlations, Gilbert, and Ros correlations. As it can be seen from Fig. 8 that the
na

match between the predicted and actual oil flow rate is poor with R2 of 0.754 using Gilbert’s

correlation. The Gilbert’s correlation underestimates the oil flow rate compared to the actual oil
ur

flow rate measured at the wellhead. On the other side, Fig.9 shows that Ros’s correlation
Jo

overestimates the predicted oil flow rate compared to the actual values. The figure also shows that

the match is weak with R2 of 0.654. in addition, statistical analysis is performed and MSE is

determined for each correlation. Fig. 10 shows the mean square error between actual and predicted

oil flow rate for the new developed correlation, Gilbert, and Ros correlations. This Fig.10 also show

that the new correlation has the highest accuracy with MSE=0.025 through achieving the highest

correlation coefficient R2 of 0.988 in comparison to existing correlations.

20
of
ro
-p
Fig. 8. Scatter diagram comparing predicted Qo and actual Qo using Gilbert’s correlation.
re
lP
na
ur
Jo

Fig. 9. Scatter diagram comparing predicted Qo and actual Qo using Gilbert’s correlation.

21
of
ro
-p
Fig. 10 MSE for the developed, Gilbert and Ros correlations
re
lP

4.2 Prediction of new oil flow rate data using ANN model
na

The developed novel correlation in this study is tested for a new dataset collected form a well #7

located in Gulf of Suez use ESP pump to lift oil with different reservoir quality. This dataset was
ur

never used in the formulation of the developed correlation. The developed correlation in this study
Jo

can predict the oil flow rate precisely as shown in Fig.11 with R2 of 0.955.

22
of
ro
-p
Fig.11. Data set used to validate the ANN model collected from Well37.
re
5. Conclusion
lP

1. The artificial neural network technique has effectively shown in prediction of oil
na

rate for artificial ESP well with an accuracy of 98-99%, as results were tested
ur

against actual wellhead data.


2. The outcome of the presented ANN in this study is the proposing a new correlation
Jo

for estimating the oil flow rate, which can apply in any case especially for wells in
Gulf of Suez Egypt having the same data range of the developed model without
using the ANN software.
3. The backpropagation learning algorithm is selected to be used combined with the
ANN technique in this study as this algorithm provides the optimum results in terms
of the lowest error and highest R2 value.
4. The developed ANN technique is tested and validated against published correlations
and gives the lowest men square error that indicated the technique is robust for oil
flow rate prediction.
5. Finally, it is recommended to test the new correlation presented in this study for
other ESP wells in different oil fields to test its strength.

23
6. References

1. Lannom DA, Hatzignatiou DG. Multiphase-flow choke correlation limits analyzed. Oil and
Gas Journal. 1996 Apr 8;94(15).
2. Nasriani, H.R. and Kalantari, A.S.L., 2011, January. Two-phase flow choke performance in
high rate gas condensate wells. In SPE Asia Pacific Oil and Gas Conference and Exhibition.
Society of Petroleum Engineers.
3. Hosseini, A.M., Ghasemzade, A. and Safian, G.A., 2011. Multiphase flow meters: technology
and applications in upstream oil and gas industry. In Presented at the 14th International Oil,
Gas and Petrochemical Congress, Tehran, Iran.
4. Folgerø, K., Lyng, S.E., Kocbach, J., Frøysa, K.E., Kleppe, K. and Åbro, E., 2013, October.
Uncertainty analysis of multiphase flow meters used for allocation measurements: Field
experiences and future challenges. In International North Sea Flow Meaurement Workshop.

of
5. Tangren, R.F., Dodge, C.H. and Seifert, H.S., 1949. Compressibility effects in two-phase
flow. Journal of Applied Physics, 20(7), pp.637-645.

ro
6. Gilbert, W.E., 1954, January. Flowing and gas-lift well performance. In Drilling and
production practice. American Petroleum Institute.
-p
7. Baxendell, P.B., 1958. Producing Wells on Casing Flow-An Analysis of Flowing Pressure
Gradients.
re
8. Ros, N.C.J., 1960. An analysis of critical simultaneous gas/liquid flow through a restriction
and its application to flowmetering. Applied Scientific Research, 9(1), p.374.
lP

9. Achong, I., 1961. Revised Bean Performance Formula for Lake Maracaibo Wells, internal co.
report. Shell Oil Co., Houston, TX.
na

10. Poettman, F.H. and Carpenter, P.G., 1952, January. The multiphase flow of gas, oil, and water
through vertical flow strings with application to the design of gas-lift installations.
In Drilling and production practice. American Petroleum Institute.
ur

11. Osman, M.E. and Dokla, M.E., 1990, January. Gas condensate flow through chokes.
In European Petroleum Conference. Society of Petroleum Engineers.
Jo

12. Perkins, T.K., 1993. Critical and subcritical flow of multiphase mixtures through chokes. SPE
Drilling & Completion, 8(04), pp.271-276.
13. Moraveji, M.K. and Naderi, M., 2016. Drilling rate of penetration prediction and optimization
using response surface methodology and bat algorithm. Journal of Natural Gas Science and
Engineering, 31, pp.829-841.
14. Nazi, G.M., Ashenayi, K., Lea, J.F. and Kemp, F., 1994. Application of artificial neural network
to pump card diagnosis. SPE Computer Applications, 6(06), pp.9-14.
15. Nygaard, R., Hareland, G., Budiningsih, Y., Terjesen, H.E. and Stene, F., 2002, September.
Eight years experience with a drilling optimization simulator in the North Sea. In SPE/IADC
Asian Conference, Jakarta, Indonesia (pp. 9-11).
16. Osborne, D.A., 1992. Neural networks provide more accurate reservoir permeability. Oil and
Gas Journal;(United States), 90(39).
17. Rumelhart, D.E., Hinton, G.E. and Williams, R.J., 1986. Learning representations by back-
propagating errors. nature, 323(6088), pp.533-536.
18. Dayhoff J.E., 1990. Neural Network Principles. Prentice-Hall Press, New York, USA

24
19. Azim, R.A., Application of artificial neural network in optimizing the drilling rate of
penetration of western desert Egyptian wells.(2) 1172 2020.
20. Mirzaei-Paiaman, A. and Salavati, S., 2013. A new empirical correlation for sonic
simultaneous flow of oil and gas through wellhead chokes for Persian oil fields. Energy
Sources, Part A: Recovery, Utilization, and Environmental Effects, 35(9), pp.817-825.
21. Gorjaei, R.G., Songolzadeh, R., Torkaman, M., Safari, M. and Zargar, G., 2015. A novel PSO-
LSSVM model for predicting liquid rate of two phase flow through wellhead chokes. Journal
of Natural Gas Science and Engineering, 24, pp.228-237.
22. Nejatian, I., Kanani, M., Arabloo, M., Bahadori, A. and Zendehboudi, S., 2014. Prediction of
natural gas flow through chokes using support vector machine algorithm. Journal of Natural
Gas Science and Engineering, 18, pp.155-163.
23. Ghareeb, M. and Elgaghah, S., 2007, January. A new correlation for calculating wellhead

of
production considering influences of temperature, GOR, and water-cut for artificially lifted
wells. In International Petroleum Technology Conference. International Petroleum

ro
Technology Conference.
24. Khorzoughi, M.B., Beiranvand, M.S. and Rasaei, M.R., 2013. Investigation of a new
-p
multiphase flow choke correlation by linear and non-linear optimization methods and
Monte Carlo sampling. Journal of Petroleum Exploration and Production Technology, 3(4),
re
pp.279-285.
25. Seidi, S. and Sayahi, T., 2015. A new correlation for prediction of sub-critical two-phase flow
lP

pressure drop through large-sized wellhead chokes. Journal of Natural Gas Science and
Engineering, 26, pp.264-278.
na

26. Leal, J., Al-Dammen, M., Villegas, R., Bolarinwa, S., Aziz, A., Azly, A., Buali, M. and Garzon, F.,
2013, March. A new analytical model to predict gas rate volume measurement through well
head chokes. In IPTC 2013: International Petroleum Technology Conference (pp. cp-350).
ur

European Association of Geoscientists & Engineers.


27. Ghorbani, H., Wood, D.A., Moghadasi, J., Choubineh, A., Abdizadeh, P. and Mohamadian, N.,
Jo

2019. Predicting liquid flow-rate performance through wellhead chokes with genetic and
solver optimizers: an oil field case study. Journal of Petroleum Exploration and Production
Technology, 9(2), pp.1355-1373.
28. AlAjmi, M.D., Alarifi, S.A. and Mahsoon, A.H., 2015, March. Improving multiphase choke
performance prediction and well production test validation using artificial intelligence: a
new milestone. In SPE digital energy conference and exhibition. Society of Petroleum
Engineers.
29. Hasanvand, M. and Berneti, S.M., 2015. Predicting oil flow rate due to multiphase flow meter
by using an artificial neural network. Energy Sources, Part A: Recovery, Utilization, and
Environmental Effects, 37(8), pp.840-845.
30. Naseri, S., Tatar, A. and Shokrollahi, A., 2016. Development of an accurate method to
prognosticate choke flow coefficients for natural gas flow through nozzle and orifice type
chokes. Flow Measurement and Instrumentation, 48, pp.1-7.
31. Choubineh, A., Ghorbani, H., Wood, D.A., Moosavi, S.R., Khalafi, E. and Sadatshojaei, E., 2017.
Improved predictions of wellhead choke liquid critical-flow rates: modelling based on
hybrid neural network training learning based optimization. Fuel, 207, pp.547-560.

25
32. Wessels, L.F. and Barnard, E., 1992. Avoiding false local minima by proper initialization of
connections. IEEE transactions on neural networks, 3(6), pp.899-905.
33. Rasoul, R.R.A. and Refaat, E., 2011, January. A Case Study: Production Management Solution"
A New Method of Back Allocation Using Downhole Pressure and Temperature
Measurements and Advance Well Monitoring". In SPE/DGS Saudi Arabia Section Technical
Symposium and Exhibition. Society of Petroleum Engineers.
34. Raikar, R. V., Wang, C. Y., Shih, H. P., & Hong, J. H. (2016). Prediction of contraction scour
using ANN and GA. Flow Measurement and Instrumentation, 50, 26-34.

of
ro
-p
re
lP
na
ur
Jo

26
1. This paper presents an intensive neural network model ANN to predict oil flow rate for
artificially flowing wells
2. The back-propagation algorithm is used as learning algorithm
3. A new correlation is developed based on the ANN model

of
ro
-p
re
lP
na
ur
Jo
The author declares that he has not any conflict of interest

of
ro
-p
re
lP
na
ur
Jo

You might also like