Oil & Gas Science and Technology - Rev.

IFP Energies nouvelles 73, 29 (2018) Available online at:

Ó D. Martinez Ricardo et al., published by IFP Energies nouvelles, 2018 www.ogst.ifpenergiesnouvelles.fr
https://doi.org/10.2516/ogst/2018031

REGULAR ARTICLE

Multiphase gas-flow model of an electrical submersible pump

Diana Marcela Martinez Ricardo1,2,*, German Efrain Castañeda Jiménez1,2, Janito Vaqueiro Ferreira1,
and Pablo Siqueira Meirelles1
1
Department of Computational Mechanics, UNICAMP-University of Campinas, São Paulo, Brazil
2
Students-Agreement Graduate - PEC-PG, CAPES/CNPq Brazil

Received: 7 December 2016 / Accepted: 2 July 2018

Abstract. Various artificial lifting systems are used in the oil and gas industry. An example is the Electrical
Submersible Pump (ESP). When the gas flow is high, ESPs usually fail prematurely because of a lack of infor-
mation about the two-phase flow during pumping operations. Here, we develop models to estimate the gas flow
in a two-phase mixture being pumped through an ESP. Using these models and experimental system response
data, the pump operating point can be controlled. The models are based on nonparametric identification using
a support vector machine learning algorithm. The learning machine’s hidden parameters are determined with a
genetic algorithm. The results obtained with each model are validated and compared in terms of estimation
error. The models are able to successfully identify the gas flow in the liquid-gas mixture transported by an ESP.

1 Introduction reflected in changes in the current during the measurement

period [1]. Another method for identifying a gas-lock is to
Different artificial lifting methods that ensure the flow of oil monitor the output flow from the pump so that the pump
through a production line are used in oil and gas production can be turned off when there is no fluid in it [3]. Monitoring
systems to reduce production losses. One such method of this kind, however, requires premature shutdown of the
involves the use of Electrical Submersible Pumps (ESPs), pump, potentially reducing well productivity. Other strate-
which were developed in 1910 by Annals Arutunoff [1]. gies that have been developed involve characterizing the
ESPs are centrifugal pumps that operate under multiphase flow by analyzing signals from measuring instruments
flow conditions and are positioned at the end of the produc- installed in the pipeline [4] and studying the fluid frequency
tion piping. They are driven by a three-phase electric motor response by measuring the vibration of the pump [5, 6].
and are suitable for use in wells with high-viscosity fluids, However, these methods can suffer from drawbacks depend-
high flow rates, high water content and a low gas-oil ratio. ing on where and how the instrumentation is installed in the
To ensure smooth operation of the pump, the gas-oil ratio pipeline.
should not exceed 10%. Exceeding this ratio can cause As the gas-lock phenomenon in this context is not well
pumping to stop because of a phenomenon known as understood and theoretical models are difficult to develop
gas-lock, a major limitation of ESPs [2]. To avoid this phe- because of the complexity of the process, experimental stud-
nomenon and consequent pump failure, the percentage of ies are required to enable a model to be developed that
gas in the ESP when it is operating under multiphase flow allows the gas flow in these systems to be estimated. Such
conditions (specifically a gas-liquid mixture) must be a model would also provide a better understanding of the
known. One way to avoid the problem is by stopping the real dynamics of the system of interest and be an important
pump or changing the pump operating conditions. Various tool for estimating and analyzing the behavior of the system
strategies for monitoring and analyzing the operation of parameters and so preventing possible failures [7].
ESPs have also been developed in an attempt to circumvent System identification is used to develop mathematical
the gas-lock problem. A traditional approach is to use models based on a data set obtained from the system of
ammeter charts, which can be used to identify operating interest. The methodology has been evolving for several
problems. By comparing an ammeter chart showing the decades [8]. In the 90s new areas of interest were considered,
equipment operating normally and another recorded when including system identification in the frequency domain,
an inspection is being carried out, the state of the equip- leading to nonparametric identification, which is currently
ment during the inspection can be determined, as this is widely used. The number of different methodologies being
used is increasing steadily.
One of these new methodologies is based on machine
* Corresponding author: [email protected] learning, which involves various computational techniques

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0),
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
2 D.M. Martinez Ricardo et al.: Oil & Gas Science and Technology - Rev. IFP Energies nouvelles 73, 29 (2018)

that aim to increase the automation and efficiency of knowl- subject to

edge acquisition processes through the use of data process-
ing. The main objective is to find the relationship between jy i  ðhw  xi i þ bÞj  e; i ¼ 1; 2;    ; m ð3Þ
the system variables (input/output) using samples acquired or
from the system [9]. One of the many applications of
machine learning is the identification of linear and nonlinear y i  ðhw  xi i þ bÞ  e;

systems, a task at which Support Vector Machines (SVMs) ð4Þ
y i  ðhw  xi i þ bÞ  e:
are known to be particularly efficient compared with other
learning methods [10]. Support Vector Machines are based The system loss function is estimated by the parameter
on the structural risk minimization principle, which origi- (e), which introduces a degree of tolerance when samples are
nated from statistical learning theory and was proposed penalized. As the linear function f(x) may not be able to fit
by Vapnik. This principle has been shown to yield better all the training data, slack variables (ni) and ðni Þ are added
results than the principle of empirical risk minimization to allow some errors subject to the following approximation
used in neural networks [11]. conditions:
In light of the increasing use of SVMs in different areas
y i  ð h w  x i i þ b Þ  e þ ni ;

of research, the new methodologies that have been devel- ð5Þ
oped in an attempt to improve the operation and perfor- y i  ðhw  xi i þ bÞ  e þ ni :
mance of SVMs and the good results obtained with these
algorithms, an SVM was used here as the basis of a method- Finally, the SVMR problem can be described by
ology for identifying the gas flow in an ESP. equation (6), where the goal is to find w and b that mini-
mize w [13]:
n
min 12 kw k2 þ C ni þ ni
P
2 Optimization of SVMR parameters with w
8
i¼1

genetic algorithms < i ð h w  x i i þ bÞ  e þ ni

> y  ð6Þ
subject to y i  ðhw  xi i þ bÞ  e þ ni
To develop a model that identifies the gas flow in an ESP
ni ; ni  0
>
:
operating with a two-phase (liquid-gas) fluid from data
acquired directly from the system, a nonparametric identi-
where C > 0 is the parameter that penalizes permissible
fication method based on machine learning was developed.
errors.
Specifically, a Support Vector Machine for Regression
If the system is not linear, the training data x 2 X
(SVMR) was used and the SVMR parameters were opti-
belonging to space X must be mapped to a higher-dimen-
mized with Genetic Algorithms (GAs).
sion space F using a function / called a kernel:
Support Vector Machines are learning systems based on
/ : X ! F: ð7Þ
optimization tools that seek to minimize structural risk.
They use a hypothesis space of linear functions in a high- The problem discussed in this paper can be represented
dimensional space created by a kernel function. These by a nonlinear system, and the kernel used to perform
transformations can be used in various learning problems, higher-dimensional mapping is a Gaussian function. The
usually either classification or regression problems [12]. parameter to be optimized with a GA is r2, which determines
Support Vector Machine for Regression aims to esti- the width of the function. The kernel
function is given by:
mate a function f(x) where the output yi will depend on jxxi j2
an input xi. Given a data approximation problem involving /ðx; xi Þ ¼ e 2 r2 : ð8Þ
a set of training data xi, yi where x 2 Ra and x 2 R, the
objective is to find a linear function f(x) that is an approx- Figure 1 shows an example of transformation of the
imation to the system using a vector of minimum weights w, mapping space by the kernel function to a larger space
where the nonlinear system behaves as if it were linear.
f ðxÞ ¼ hw  xi þ b Equation (6) can be solved using dual programming
[14]. The goal is to build an objective function by adding
¼ xT w þ b a set of variables ai called Lagrange multipliers, where the
n ð1Þ
¼
P
wi xi þ b dual formulation is given by
i¼1 "
n
 

max  12 ai  ai aj  aj / xi ; xj
P
where the training data sets are x = x1,x2,  , xn and i;j¼1
y = y1,y2,  ,yn, the weight vector is w = w1,w2,  ,wn n n

and b is the bias. e
P
ai þ ai


þ y i ai þ
P
ai


To achieve this objective, the vector w must be mini- i¼1 i¼1
: ð9Þ
mized using the Euclidean norm ||w||2. The problem can 8 n
< P a  a
¼ 0
then be defined as the optimization problem i i
subject to i¼1
1 2
:
ai ; ai 2 ½0; C 
min jjwjj ð2Þ
w 2
 D.M. Martinez Ricardo et al.: Oil & Gas Science and Technology - Rev. IFP Energies nouvelles 73, 29 (2018) 3

7
Gaussian

6
Kernel

5
4
3
2
1

1 2 3 4 5 6 7 8

Fig. 1. Mapping the vectors into higher-dimensional space.

After solving the dual problem the optimal decision is

given by
n
X
ai  ai / xi ; xj þ b:



f ðxÞ ¼ ð10Þ
i¼1

Correct selection of the hidden parameters of an SVM

ensures that the models developed using the SVM perform
satisfactorily. The parameters are C, which determines the
curvature of the margin penalizing permissible errors, e,
which defines the insensitivity of the margin, and r2, which
determines the width of the kernel function. In this article,
the use of a GA is proposed to search for the best hidden
parameters of the SVMR.
Genetic Algorithms provide a robust method for per-
forming searches and optimization. They mimic the evolu-
tionary processes of living things and can provide
solutions for a wide range of problems. Because of their
adaptability, GAs have been used by many researchers to
look for the hidden parameters of SVMRs for various prob-
lems [15, 16]. The flowchart of the GA used in this paper is
shown in Figure 2.
To find the best parameters (C, e,r2), the algorithm
must perform the following four steps. The first step con-
sists of creating a population of individual elements
pk ¼ xt1    xtn or pk ¼ ðC; ; r2 Þt1    ðC; ; r2 Þtn that are

candidates for a possible solution. In the second step, the
adaptation of each of the individual elements to the prob-
lem being studied is measured using a fitness function. In
this particular case, the validation error of each of the mod-
els built with the SVMR is used as the fitness criterion, and
the objective function is given by
1
Fitness ¼ max eval
nval
2 ð11Þ Fig. 2. Flowchart of the genetic algorithm.
) eval ¼ 2n1val
P
f ðxval Þi  y val
i¼1
be selected, helping to maintain the diversity of the popula-
where nval is the number of validation samples. tion of solutions and so avoid convergence to local minima.
In the third step, the individuals that are better adapted In the fourth step, a crossover is performed between two
to the problem are selected. This is done by the roulette or more individuals selected in the previous stage (parents)
method, which not only ensures a higher probability of bet- to generate descendants with part of the genetic material
ter-adapted individuals being selected, but also also allows from each parent. For the crossover to be performed, all
individuals that are less well adapted to the problem to individuals must be coded in binary representation using
4 D.M. Martinez Ricardo et al.: Oil & Gas Science and Technology - Rev. IFP Energies nouvelles 73, 29 (2018)

the same number of bits for each parameter [e.g., Table 1. ESP test-circuit configuration for data
P = (C, e, r2) = (2,4,3) ! Pbit = (0010 0100 0011)]. acquisition.
Hence, if a parameter is represented by b bits, the genome
size is l = 3b. In the crossover process, a random point k is Test Pump speed Inlet pressure Gas flow
selected and the bits to the right of this are swapped (rpm) (bar) (kg h1)
between the parents to create descendants [e.g., if the par- 1 1800 100 0
ents are P1 = (0010 0100 0011) and P2 = (0101 0001
1
1111), then for k = 5 we have S1 = (0010 0001 1111) and
S2 = (0101 0100 0011) as descendants]. This crossover pro- 200 2
cess is repeated until the population of descendants is the 3
same size as the initial population. The algorithm includes 2 3000 100 0
a final step called the mutation operator that is intended 1
to maintain population diversity. This operator randomly
200 2
selects some of the individuals generated and performs a
mutation in their genomes by changing one bit at random 3
[e.g., a selected individual S = (0010 0001 1111) becomes
Smutate = (0010 0001 1011) after the mutation]. The above
steps are repeated until the stop condition is satisfied.
models were tested. The different sets of training samples
used for each model are shown in Table 2. The objective
of these 10 sets is to be able to select as input the best com-
3 Results bination of the data acquired from the test system to be
able to identify the gas flow in the EPS and thus obtain
An SVMR was used to develop a nonparametric black-box models that represent the variations of the percentage of
model for identifying the gas flow in a J200N ESP based on gas in the pump.
different characteristic system parameters. Casing vibra- Having defined and acquired the model inputs and out-
tion, total flow, torque and elevation were used as inputs, puts, we used a genetic algorithm to estimate the best
and a graphical representation of estimated gas flow in parameters (C, e, r2) and so identify the most representa-
the ESP was used as output, since these are the data that tive model for each set of training samples. Various tests
are available in the ESP circuit tests. Several experiments were performed for each of the ten models in Table 2.
were performed in the ESP test circuit in the LABPETRO The configuration of the GA for each of the tests is shown
laboratory at Unicamp using the initial operating condi- in Table 3.
tions in Table 1. The ESP system was tested at 1800 rpm The results of the tests with each of the models using the
and 3000 rpm with manometric inlet pressures of 100 bar GA configurations in Table 3 are shown in Table 4. Selec-
and 200 bar, respectively, and gas flows of 0–3 kg h1. tion of the parameters to be used with the SVMR was based
After the data used to estimate the model parameters on analysis of the Mean Square Error (MSE) and autocor-
had been acquired, the signals were post-processed. This relation coefficient. The parameters that yielded the lowest
is necessary because the acquired ESP casing vibration sig- MSE and highest autocorrelation coefficient were selected
nals contain noise and are difficult to predict in the time and then used in the model to predict the ESP gas flow.
domain even though the initial test-circuit conditions are Table 4 shows the configurations of the GA that yielded
known. A spectral representation of the signals was there- the best results and the corresponding SVMR parameters,
fore used so that they could be characterized by their power MSE and autocorrelation coefficient for each model.
spectral density. This was done with a finite Fourier trans- Once the input and output data for each model have
form and the autocorrelation function, a technique typically been acquired and the SVMR parameters have been esti-
used to analyze measurement signals when working with mated by the GA, the gas flow in the ESP can be estimated
digital equipment and discrete algorithms, as in this case using different characteristic system parameters. The power
[17]. of generalization of each model was determined by measur-
After the spectral representation had been generated, ing the performance of the model with a data set not
the autocorrelation of the power spectral density was calcu- included in the training data used to build the model. This
lated for specific positive frequencies corresponding to the test data set corresponded to 20% of the total data used and
pump speeds and their respective multiples. For example, the rest of the data is used for training [18].
for a pump speed of 1800 rpm, the frequencies evaluated To evaluate the ten SVMR models, the actual gas flow
were 30, 60, 90 and 120 Hz, as these are the frequencies was compared with the gas flow predicted by each model.
at which there are significant changes in the spectral repre- For each of the four flows, a mean was estimated for all
sentation when the gas flow in the ESP changes. In this the test samples and the MSE was calculated for each
way, a new set of data is acquired that can be used as inputs model so that their performance could be compared. Table 5
in the system identification. The patterns of the spectral shows the gas flows estimated by the SVMR models and the
representations of the signals associated with the casing corresponding MSE.
vibration variable can be seen in Figure 3. To evaluate the performance of the models, the gas
Ten different models were built using a variety of signals flows estimated by each model were compared. Figure 4
as input, and once the parameters had been estimated the shows the flows estimated by each of the models for actual
 D.M. Martinez Ricardo et al.: Oil & Gas Science and Technology - Rev. IFP Energies nouvelles 73, 29 (2018) 5

Fig. 3. Patterns of the spectral representation of casing vibration.

Table 2. Input and output data used with each model. Table 4. Selected SVMR parameters.
Model Input Output SVMR parameters
1 Spectral representation of the ESP Gas flow Model GAca C r2 e MSEb rc
vibration signal and total flow 1 7 20.698 0.127 0.0011 0.0183 0.98
2 Spectral representation of the ESP Gas flow 2 7 15.934 0.125 0.0022 0.0844 0.93
vibration signal and torque 3 3 473.793 0.345 0.0745 0.1023 0.89
3 Spectral representation of the ESP Gas flow 4 2 547.638 30.990 0.0078 0.0761 0.96
vibration signal and elevation 5 4 14.322 1.667 1.6673 0.0027 0.99
4 Total flow and torque Gas flow 6 2 55.237 2.193 0.0170 0.0582 0.95
5 Total flow and elevation Gas flow 7 4 4.061 0.463 0.0924 0.0227 0.98
6 Torque and elevation Gas flow 8 4 8.756 30.857 0.2394 0.9309 0.59
7 Spectral representation of the ESP Gas flow 9 6 777.785 31.495 0.2435 0.1610 0.83
vibration signal 10 7 903.983 1.226 0.2381 0.4413 0.71
8 Total flowl Gas flow a
9 Torque Gas flow GA configuration.
b
Training MSE.
10 Elevation Gas flow c
Autocorrelation coefficient.

Table 3. Configuration of the genetic algorithm in each

test.
flows of 0–3 kg h1. These results show that all the models
Config Itera Indb Genesc Pcrossd Pmute have a good power of generalization but that the model
1 100 10 10 50 10 with the best power of generalization is model 5, which is
2 500 10 10 50 10 based on total flow and elevation. The gas flow estimated
with this model has the lowest error of all the models.
3 1000 10 10 50 10
The model chosen to predict the gas flow in the ESP
4 100 100 10 40 20 was the model with the lowest MSE and highest autocorre-
5 500 100 10 40 20 lation coefficient.
6 1000 10 10 40 20 The results in Table 6 show that model 5, which uses
7 1000 100 10 60 10 total flow and elevation as input parameters, yielded the
8 50 10 10 60 10 best results (MSE = 0.003 and autocorrelation coeffi-
cient = 0.997) and that the experimental data provided
a
No of iterations. good information about changes in the gas flow in the
b
No of individuals. pump, as the errors in the predicted values are small for
c
No of genes. most of the models. The findings also show that when the
d
Probability of crossing. training data is of poor quality, the models do not
e produce good estimates.
Probability of mutation.
6 D.M. Martinez Ricardo et al.: Oil & Gas Science and Technology - Rev. IFP Energies nouvelles 73, 29 (2018)

Table 5. Gas flow estimated by each SVMR model.

Gas flow
Actual estimated by
Model gas flow the SVMR model MSE
1 0 0.0414 0.0195
1 0.8545 0.0370
2 1.9475 0.0115
3 3.0507 0.0028
2 0 0.0659 0.0109
1 1.2977 0.2408
2 1.8109 0.0388
3 2.9372 0.0142
3 0 0.1596 0.0395
1 1.1705 0.1425
2 1.6571 0.1714
3 3.0024 0.0001
4 0 0.2242 0.1019
1 1.2975 0.1124
2 2.1347 0.0409
3 2.9210 0.0062
5 0 0.0010 0.0008
1 0.9822 0.0059
2 2.0020 0.0006
3 2.9639 0.0020
6 0 0.0910 0.0658
1 0.7639 0.0994
2 1.7985 0.0541
3 3.0206 0.0009
7 0 0.1486 0.0226
1 1.0421 0.0474
2 1.9836 0.0002
3 2.9975 0.0069
8 0 0.6648 1.1373
1 1.0858 0.2538
2 3.2097 1.9340
3 2.7668 0.054
9 0 0.2432 0.0591
1 1.1819 0.2356
2 1.8475 0.209
3 2.7749 0.0513
10 0 0.3933 0.2680
1 1.5125 0.4626
2 1.8016 0.1280
3 2.2531 0.7857

Fig. 4. Gas flow estimates for each model.

 D.M. Martinez Ricardo et al.: Oil & Gas Science and Technology - Rev. IFP Energies nouvelles 73, 29 (2018) 7

Table 6. MSE and autocorrelation coefficient of the 3 Snyder J.R., Dale R., Joe Haws H. (1989) Pump off/gas lock
SVMR models. motor controller for electrical submersible pumps, EP Patent
App. EP19,880,202,383, 03 May.
Model MSE Autocorrelation coefficient 4 Klein F.L., Seleghim, Jr. P., Eric H. (2004) Time-frequency
02 analysis of intermittent two-phase flows in horizontal piping,
1 1.84E 0.985 J. Braz. Soc. Mech. Sci. Eng. 26, 2, 174–179.
2 6.16E02 0.954 5 Kolpak M.K., Rock T.J. (1996) Measuring vibration of a fluid
3 0.102 0.896 stream to determine gas fraction, U.S. Patent No. 5,524,475,
4 0.076 0.967 11 June.
6 Henry M.P., Richard R.P. (2014) Multiphase flow metering
5 0.0027 0.997 system, U.S. Patent Application No 14/135,085, 19 Dec.
6 0.058 0.957 7 Nelles O. (2001) Nonlinear system identification, Springer-
7 0.022 0.983 Verlag, Berlin Heidelberg.
8 0.930 0.597 8 Ljung L. (1999) System identification: Theory for the user,
Prentice Hall PTR, Upper Saddle River, NJ.
9 0.161 0.830
9 Dietterich T.G. (1986) Learning at the knowledge level,
10 0.441 0.710 Mach. Lear. 1, 287–315.
10 Rychetsky M. (2001) Algorithms and architectures for
machine learning based on regularized neural networks and
4 Conclusion support vector approaches, Shaker Verlag GmbH, Germany.
11 Vapnik V. (1995) The nature of statistical learning theory,
Springer-Verlag, New York.
We have proposed different nonparametric VSMR models 12 Vapnik V., Golowich S., Steven E., Smola A. (1996) Support
for estimating the gas flow in an ESP. The models use vector method for function approximation, regression esti-
experimental vibration, total-flow, torque and elevation mation, and signal processing, Advances in Neural Informa-
data collected from the pump and pipe system. tion Processing Systems 9, pp. 281–287.
The hidden parameters of the SVMR were estimated 13 Stoean R., Dumitrescu D., Preuss M., Stoean C. (2006)
using a GA, thus ensuring that the models have good power Evolutionary support vector regression machines, in:
of generalization and, in turn, perform well. This was con- SYNASC ’06 Proceedings of the Eighth International Sym-
firmed by testing the models with data that were not used posium on Symbolic and Numeric Algorithms for Scientific
during training. The models can be used to solve the Computing, IEEE Computer Society, Washington, DC,
problem of premature failure due to high gas flow in ESP pp. 330–335.
systems. This problem arises because of insufficient infor- 14 Zhang H., Yongmei L. (2013) BSP-Based Support Vector
mation about the multiphase nature of the flow in the Regression Machine Parallel Framework, Computer and
pump and can be easily avoided just by changing the pump Information Science (ICIS), 2013 IEEE/ACIS 12th Interna-
operating conditions. tional Conference on. IEEE, Niigata, Japan, pp. 329–334.
One of the most significant findings of this study is that 15 Yu Q., Liu Y., Rao F. (2009) Parameter selection of support
total flow and elevation yielded the most accurate estimates vector regression machine based on differential evolution
of gas flow, as the models using these parameters had the algorithm. Fuzzy systems and knowledge discovery, de FSKD
’09. Sixth International Conference, pp. 596–598.
lowest MSE (0.003) and highest autocorrelation coefficient
16 Zhenyue H., Mei C. (2009) Soft sensor modeling using SVR
(0.997).
based on genetic algorithm and akaike information criterion,
intelligent human-machine systems and cybernetics, Intelli-
gent Human-Machine Systems and Cybernetics, 2009.
References IHMSC’09. International Conference on, Vol. 2, IEEE.
17 Blackman R.B., Tukey J.W. (1959) The measurement of
1 Tákacs G. (2009) Electrical submersible pumps manual: power spectra, Dover publications, New York.
design, operations, and maintenance, Gulf Professional 18 Michael K. (1996) A bound on the error of cross validation
Publishing. using the approximation and estimation rates, with conse-
2 Varon M.P. (2013) Estudo de uma bomba centrı́fuga quences for the training-test split, Advances in Neural
submersa (BCS) como medidor de vazão, Msc Thesis, Information Processing Systems, pp. 183–189.
FEM-UNICAMP, Campinas, Brazil.