Advances in Engineering Research (AER), volume 105

3rd Annual International Conference on Mechanics and Mechanical Engineering (MME 2016)

An Artificial Intelligence Prediction Method of Bottomhole Flowing

Pressure for Gas Wells Based on Support Vector Machine
Qin-Feng DI1,a, Wei CHEN1, Jing-Nan ZHANG1, Wen-Chang WANG2,a and
Hui-Juan CHEN2
1 Shanghai Institute of Applied Mathematics and Mechanics, Shanghai University, Shanghai 200072,
China
2 Shanghai Key Laboratory of Mechanics in Energy Engineering, Shanghai 200072, China

Keywords: Flowing bottomhole pressure, Support vector machine, Random samples selection, Gas
wells.

Abstract. The flowing bottomhole pressure (FBHP) of gas wells was affected by many factors.
Although a lot of research works have been done to predict the FBHP and at least more than ten
models were proposed, but no one can effectively provide an accurate results for all ranges of
production data and conditions due to the existence of many uncertain relations between the
changeable influence factors. In this paper, an artificial intelligence prediction method for FBHP
based on the support vector machine (SVM), named the FBHP-SVM method, was studied, and a
support vector regression (SVR) model with ε-insensitive loss function (ε-SVR) based on radial basis
function (RBF) was used to predict the FBHP of gas wells. Compared with the true values, the
average absolute and relative errors of the new method were 0.27MPa and 2.29%, respectively. The
FBHP-SVM method was also compared to the vertical pipe flowing method. The results showed this
new method was a new practical tool to predict FBHP in gas wells and it had a satisfying prediction
accuracy.

Introduction
Accurately predicting the flowing bottomhole pressure (FBHP) of gas wells is the basis of dynamic
analysis and production strategies optimization. The FBHP prediction is very complicated because
many parameters have influence on the FBHP in the wellbore and they are continuous changing in
the process of production. Although the first empirical formula has been put forward for many
decades and many new methods have been proposed continually, there is no an applicable available
model due to the existence of many uncertain relations between the changeable influence factors [1].
These prediction methods can be roughly classified into three kinds, i.e. empirical correlations,
mechanistic models and artificial intelligence methods.
Many empirical correlations for FBHP prediction have been developed since the early 1940s.
Most of these correlations were proposed by the investigators from laboratory studies, including
those of Duns and Ros [2], Hagedorn and Brown [3], Beggs and Brill [4] and Orkiszewski [5].
Generally, these empirical correlations had a perfect performance under the condition that the model
was proposed. But when the prediction conditions differ from the specific boundaries that the model
was proposed, the FBHP prediction accuracy tends to decrease. Under this circumstances, the
mechanistic models were developed to predict FBHP for gas wells. Most of them were semi-
empirical models and had sound theoretical foundation and wider application than empirical
correlations [6]. The widely used mechanistic models are those of Rzasa and Katz [7], Cullender and
Smiths [8], Ansari et al. [9], Chokshi et al. [10], Hasan and Kabir [11]. However, some empiricism,
more or less, are still involved to overcome the complexity of the problem and the mechanistic
models are very difficult to meet the requirement of a complicated well in which the continuous
changes of the temperature and pressure will directly affect the gas production rate. And the gradual
change in gas volumes will lead to the change of liquid slip velocity and the appearance of new flow
patterns. The variation of flow patterns and their transition boundaries inside the well bore

Copyright © 2017, the Authors. Published by Atlantis Press. 206

This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).
 Advances in Engineering Research (AER), volume 105

contributed to the deviation in FBHP predictions. In this condition, in order to decrease the
deviations, the most suitable mechanistic model for each of these flow patterns needs to be used to
predict the FBHP, which adds the complexity of calculation process. Thus, a method that can be
applied to all ranges of production and conditions is urgent to be studied, and without doubt, new
approaches that go beyond the conventional analysis methods are good choices.
In the recent few decades, the artificial intelligence methods are frequently used in oil and gas
industries, and their unique advantages of solving various difficult engineering problems have drawn
a great deal of attention. To overcome above-mentioned shortcomings associated with the empirical
correlations and mechanistic models, researchers made use of artificial intelligence-based methods
foremost of which was the classical artificial neural network (ANN) and its variants. But still, ANN
had several inherent drawbacks such as over fitting, slow convergence, poor generalizing
performance, and arriving at local minimum, and so on [12]. In order to achieve this target, a new
type of artificial intelligence tool based on support vector machine (SVM) was utilized to make
FBHP prediction in this paper. SVM is a new type of machine learning algorithm. It predicts
unknown data by self-learning of the known data and can obtain the inherent laws even if it cannot
be gotten by theory analysis. In addition, the SVM method has many unique advantages, such as
strong generalization and convergence to global optimization, and dimensional insensitivity [13]. All
of these characteristics make it possible to achieve the FBHP prediction no matter how complicated
the conditions are, even in directional wells or horizontal wells. The works in this paper focus on
vertical wells, and include four aspects. Firstly, a model based on radial basis function was used to
establish the FBHP prediction model of gas wells and a FBHP prediction program was developed
with MATLAB. Secondly, the effectiveness of the model and the program were tested by an
exponential function with two independent variables and a noise item. Thirdly, the field data
collected from Sulige gas field were used to evaluate the prediction accuracy based on random
selection of the verification sample. Finally, the reliability of new developed method was verified by
comparing with an existing widely used mechanistic model.

Methodology of Support Vector Method

SVM was proposed by Vapnik in 1995 based on the structural risk minimization principle and was
an effective machine learning method for classification problems, especially for small sample of
training vectors [14]. It studies how to construct learning machine, to realize the pattern classification.
It aims to seek an optimal hyperplane that correctly classifies data points as much as possible and
separates the points of two classes as far as possible [15]. SVM is considered as a highly efficient
better performance classifier, so that it can replace the multi-layer perceptron, radial basis function
(RBF) neural network and the polynomial network and so on other existing learning algorithm [16].
Support Vector Regression
The support vector regression (SVR) is the application of the SVM on regression problems and is
based on the theory of linear regression in a multidimensional feature space [17]. The optimal
hyperplane of SVR is to make the overall deviation between the predicted values and the true values
minimum [18].
In the linearly separable case, Fig.1 shows the example of one-dimensional linear regression
function with ε insensitive region. The variable ξ measures the error of the training point and the
error is zero in ε insensitive region. In the nonlinearly separable case, the input data x can be
transformed into the higher dimensional feature space where x is linearly separable by using a kernel
function. The support vector regression model with ε insensitive region is named ε-SVR [19].

207
 Advances in Engineering Research (AER), volume 105

Figure 1. One-dimensional linear regression function with ε insensitive region [19].

Assuming a dataset T  {( xi , yi ), i  1,..., n} , wherein the independent variables are xi  R d (d -

dimensionality) and the related dependent variables are yi  R , an objective function f(x) needs to be
found to predict the corresponding function values of any independent variables. Because the ε-SVR
model can be converted into a dual problem of a convex quadratic program optimization, the dual
problem can be represented as the maximizing function of Eq. (1) under the constraints of Eq. (2)
[20].
(1)
1 n
 i   i*  j   *j  K  xi , x j     i   i*  yi    i   i*  
n n
W ( ,  * )   
2 i , j 1 i 1 i 1

where α and α* stand for the Lagrange multipliers; n is the sample size; i and j represent the
sample numbers and K ( xi , x j ) stands for the kernel function [20].

  0
n

 i
*
i
(2)
i 1

0   i  C , 0   i  C , i  1, 2,..., n
*

where C stands for the penalty factor. The objective function can be expressed as [20]:

  
  i K  xi , x j   b
n

f ( x) 
*
i
(3)
i 1

where b is the classification threshold and can be obtained by the following equations [20]:
  yi  f  xi   0,    0, C 
 , i  1, 2,..., n (4)
  yi  f  xi   0,    0, C 
*

Solving  -SVR Model by Self-developed Program

The SVM models can be solved as a quadratic program problem [18]. A self-developed program
based on the ε-SVR model was used to predict the FBHP of gas wells, which is named as the FBHP-
SVM method in this paper. The block diagram was shown in Fig.2. In the prediction model, three
parameters of the kernel parameter σ, insensitive factor ε and penalty factor C have significant
effects on prediction results. After the initial ranges of the three parameters were determined, the web
search method was used for parameters optimization. The parameters that accords with the minimum
training sample mean square error (MSE) were selected to establish the model.
n
1
 f  x   y 
2
min i i
(5)
 ,C , n i 1

where f (xi) and yi are the predicted values and the true values, respectively.

208
 Advances in Engineering Research (AER), volume 105

Figure 2. MATLAB program block diagram of the ε-SVR model.

Method Verification
A validity test must be performed in order to evaluate the new method. An exponential function with
two independent variables and a noise item was used to test the effectiveness of the self-developed
program. Assuming a test function z  10e x  y   , wherein ζ is a normal distribution noise item and
2 2

 ~ N  0, 0.01 . In x  [1,1] and y  [1,1] , 121 points were evenly selected to form the data set when
the step-size is 0.2, wherein 113 points were randomly picked as the training samples (TRS) and the
remaining samples were as the verification samples (VES). The data set collected were showed in
Fig.3. The initial ranges of three parameters were given as follows: σ is between 0.1 and 10 and the
step is 0.2; ε is between 0.1 and 3 and the step is 0.1; C is between 10 and 100 and the step is 10.
Inputting data into the self-developed program and optimizing the model parameters, a set of optimal
parameters of σ, ε and C was determined, i.e. 1.1, 0.1 and 10, respectively, and the corresponding
MSE of training samples is 0.2217. The comparison results between true values and prediction
results of the FBHP-SVM method were shown in Fig.4 and the detailed data of verification samples
were listed in Table 1.

Figure 3. Division of samples collected.

Figure 4. Comparison between the true values and predicted values of the FBHP-SVM method.

209
 Advances in Engineering Research (AER), volume 105

Table 1. Comparison between the true values and predicted values with verification samples.
z
x y
True values Predicted values Absolute errors /MPa Relative errors /%
-1 0.2 3.544 3.428 0.116 3.28
-0.8 0.2 5.072 5.060 0.012 0.24
-0.6 0.6 4.869 4.806 0.063 1.29
-0.4 0.2 8.192 8.291 0.099 1.21
0.2 0.2 9.240 9.432 0.192 2.08
0.4 1 3.137 3.312 0.175 5.58
0.8 0.6 3.688 3.766 0.078 2.12
1 0.2 3.544 3.820 0.275 7.76

Figure 4 shows that the differences between the predicted values with training samples (PV-TRS)
and the true values of training samples (TV-TRS) are small. From Table 1, it is clearly observed that
the largest absolute and relative errors between the predicted values with verification samples (PV-
VES) and the true values of verification samples (TV-VES) are 0.275 MPa and 7.76%, respectively.
It illustrates that the self-developed program can be applied to predict FBHP of gas wells.

Data Collection and Pre-processing

Data Collection
Sixteen groups of samples obtained from well testing were collected from Sulige gas field of China
[21]. For an model, properly selecting the inputs is very important. Unnecessary and uncertain
inputs may increase prediction errors. After studying the different empirical correlations and
mechanistic models, the following parameters are selected as input data to predict FBHP, i.e.
wellhead pressure, Pc, well depth, H, relative density of gas,  g , average temperature in the wellbore,
T , gas compression factor, Z and gas rate, qsc.
Data Normalization
In order to avoid the magnitude difference between the input parameters and eliminate the dimension
influence of different parameters, some data pre-processing must be performed in order to train the
model rapidly and with more accuracy. The input data are needed to be transformed into a unified
form with normalization. The normalization formula is as follows:
2( xij  x j min )
xij  1
'
(6)
x j max  x j min

where xij and xij’ are the jth influence factors of the ith training sample before and after
normalization; xjmax and xjmin are the maximum and minimum values of the jth influence factor in all
the training samples.

FBHP Prediction of Gas Wells Based on SVM

After the influence factors and the value ranges of σ, ε and C were determined, the ε-SVR model for
FBHP prediction of gas wells was established. Five groups of samples, in turn, were randomly
picked as the verification samples. For every group of verification sample, a model was firstly
established by importing the remaining fifteen groups of samples into the self-developed program as
training samples according to the principal of minimum training sample mean square error, then the
prediction value of FBHP was obtained by inputting the wellhead pressure, well depth, relative
density of gas, average temperature in the wellbore, gas compression factor and gas rate of the
verification sample into this model. The results were shown in Table 2 and Fig. 5.

210
 Advances in Engineering Research (AER), volume 105

Table 2 and Fig. 5 have shown a good concordance of the predicted values of five groups of
verification samples obtained by the FBHP-SVM method (PV-SVM) with the true values of
verification samples (TV-VES).The average absolute and relative errors of the five groups of
verification samples are 0.27MPa and 2.29%, respectively, which indicates that the FBHP-SVM
method proposed in this paper can be used to predict the FBHP of gas wells with a satisfying
accuracy.

Figure 5. Comparison between the true values and predicted values of the FBHP-SVM method.

Table 2. Comparison between the true values and predicted values of the FBHP-SVM method
Pwf/MPa
Number

qsc /
Relative
Ln(Pc) g T Absolute
H/m Z True Predicted errors
/MPa / °C 104m3/d values values
errors
/%
/MPa
1 1.96 3273.4 0.6029 67.15 1.531 0.9873 8.82 8.15 0.67 7.60
2 2.35 3292.9 0.6065 63.90 1.126 0.9942 13.13 13.24 0.11 0.84
3 2.65 3331.7 0.6066 64.00 2.498 0.9873 17.38 17.54 0.16 0.92
4 2.67 3335.0 0.6187 55.75 2.502 0.9869 18.12 18.19 0.07 0.39
5 2.71 3312.9 0.6146 70.95 3.121 0.9915 19.69 20.02 0.33 1.68
Average error 0.27 2.29

Compared with the Vertical Pipe Flowing Method

To demonstrate the superiority of the method proposed in this paper, the predicted results were
compared with those predicted by the vertical pipe flowing method [22]. As a simplified mechanistic
model, the vertical pipe flowing method’s advantages are that it requires fewer required parameters
and has a simple calculation process. Currently, vertical pipe flowing method is widely applied in the
field.
Vertical Pipe Flowing Method
According to the pressure gradient equation in vertical gas wells, the expression for calculating
FBHP by the vertical pipe flowing method under the condition of ignoring the kinetic pressure
gradient is as follows [23]:
18
pwf  pc e  1.32  10 f (T Zqsc ) (e  1) / D (7)
2 2s 2 2s 5

where Pwf is the FBHP, MPa; Pc is the casing pressure, MPa; f is the friction resistance, non-
dimensional; T is the average temperature in the wellbore, K; Z is the gas compression factor, non-
dimensional; qsc is the gas rate, m3/d; and D is the tube diameter, m. s is the exponent [23].

211
 Advances in Engineering Research (AER), volume 105

0.03418 g H (8)
s
TZ
where g is the relative density of gas, non-dimensional; and H is the well depth, m.
Because the friction factor f varies with pipe surface roughness which is related to pipe material,
manufacturing technology, corrosion and scaling condition, so it is very difficult to obtain accurate
friction factor values. According to Wanglin’s results, the friction factor can be ignored [17]. Then
Eq. (7) can be written as follows:
0.03415  g H

pwf  pc e TZ
(9)

Five groups of verification samples were used to calculate FBHP by the vertical pipe flowing
method. The results are compared with the true values in Table 3 and Fig. 6.
From Table 3 and Fig. 6, it is clear that the average absolute and relative errors between the true
values and predicted results with vertical pipe flowing method (PV-VP) are 0.40 MPa and 2.39%,
respectively, which are in the acceptable range and can meet the needs of practical production.

Figure 6. Comparison between true values and predicted values of vertical pipe flowing method.

Table 3. Comparison between the true values and predicted values of vertical pipe flowing method

Pwf /MPa Absolute Relative

PC / H g T
Number Z True Predicted errors errors
MPa /m /K
values values /MPa /%
1 7.10 3273.4 0.6029 340.15 0.9873 8.82 8.68 0.14 1.59
2 10.50 3292.9 0.6065 336.9 0.9942 13.13 12.87 0.26 1.98
3 14.20 3331.7 0.6066 337 0.9873 17.38 17.47 0.09 0.52
4 14.40 3335 0.6187 328.75 0.9869 18.12 17.89 0.23 1.27
5 15.00 3312.9 0.6146 343.95 0.9915 19.69 18.39 1.30 6.60
Average error 0.40 2.39
Results Comparison and Analysis
The comparison between the results obtained by the FBHP-SVM method and vertical pipe flowing
method are shown in Table 4 and Fig. 8.
As Table 5 and Fig. 7 show, the average absolute errors between the true values and predicted
results of the FBHP-SVM method and the vertical pipe flowing method, are 0.27 MPa and 0.40 MPa
and the average relative errors are 2.29% and 2.39%, respectively, both of which indicate that the
FBHP-SVM method can be used as a new practical way to predict FBHP with satisfying accuracy.

212
 Advances in Engineering Research (AER), volume 105

Especially, the prediction results of the vertical pipe flowing method are affected by the
parameters in the equation (7) ~ (9). When the pressure and temperature of formation change, the
calculation formula or the value ranges of the friction resistance, f , the relative density of gas,  and g

the gas compression factor, Z will be different. Moreover, for directional wells or horizontal wells,
the calculation method of the friction resistance, f is also different from the vertical wells. Thus, for a
complicated well, the prediction for FBHP calls for different methods for different regions of flow
patterns and different deviation angles. The proposed FBHP-SVM method in this paper is applicable
for all ranges of production conditions, so the FBHP-SVM method is a more practical tool to predict
FBHP for gas wells.

Table 4. Comparison of results

Pwf /MPa Absolute errors/MPa Relative errors/%
Samples FBHP- Vertical pipe Vertical pipe FBHP- Vertical pipe
True FBHP-SVM
SVM flowing flowing SVM flowing
values method
method method method method method
1 8.82 8.15 8.68 0.67 0.14 7.60 1.59
2 13.13 13.24 12.87 0.11 0.26 0.84 1.98
3 17.38 17.54 17.47 0.16 0.09 0.92 0.52
4 18.12 18.19 17.89 0.07 0.23 0.39 1.27
5 19.69 20.02 18.39 0.33 1.30 1.68 6.60
Average error 0.27 0.40 2.29 2.39

Figure 7. Results comparison.

Conclusions
The FBHP-SVM method was proposed in this paper to predict the FBHP of gas wells. The main
conclusions are as follows.
(1) SVM method had been successfully used to predict the FBHP in this paper and a better
performance had been gotten because it can suit all ranges of production conditions.
(2) In order to avoid human intervention, a random selection method of verification samples was
used in the process of predicting FBHP of gas wells by the FBHP-SVM method. The results showed
that the absolute and relative errors were 0.27MPa and 2.29%, respectively, which indicated that this
method can be used to predict flowing bottom hole pressure with a good prediction precision.
(3) The results showed that the FBHP-SVM method had an approximate accuracy compared with
the vertical pipe flowing method. It is worth noting that this new method would have better
application prospects because it does not rely on a given mechanistic model with changeable
parameters and empirical factors.

213
 Advances in Engineering Research (AER), volume 105

Acknowledgement
This research was supported in part by the National Science Funding of China (U1663205,
50874071), the Key Program of Science and Technology Commission of Shanghai Municipality
(071605102), Shanghai Leading Academic Discipline Project (S30106), and the Shanghai Municipal
Education Commission (Peak Discipline Construction Program).

References
1. Al-Shammari A. SPE Saudi Arabia Section Technical Symposium and Exhibition (Society of
Petroleum Engineers, Al-Khobar, 2011)
2. Duns Jr H, Ros N C J. 6th World Petroleum Congress (Frankfurt am Main, 1963)
3. Hagedorn A R, Brown K E. Journal of Petroleum Technology. 17, 475 (1965)
4. Beggs D H, Brill J P. Journal of Petroleum technology. 25, 607 (1973)
5. Orkiszewski J. Journal of Petroleum Technology. 19, 829 (1967)
6. Osman E S A, Ayoub M A, Aggour M A. SPE Middle East Oil and Gas Show and Conference
(Society of Petroleum Engineers, Kingdom of Bahrain, 2005)
7. Rzasa M J, Katz D L. Transactions of the AIME. 160, 100 (1945)
8. Cullender M H, Smith R V. Trans. AIME. 207, 281 (1965)
9. Ansari A M, Sylvester N D, Shoham O, et al. SPE Annual Technical Conference and Exhibition
(Society of Petroleum Engineers, New Orleans, 1990)
10. Chokshi R N, Schmidt Z, Doty D R. SPE Western Regional Meeting (Society of Petroleum
Engineers, Anchorage, 1996)
11. Hasan A R, Kabir C S. SPE. Production Engineering. 3, 263 (1988)
12. Xue X H, Yang X G, Chen X. Sci China Tech Sci. 57, 2379 (2014)
13. Zhang X G. Acta Automatica Sinica. 26, 32 (2000)
14. Liu H Q, Chen Q H, Li B, et al. Sci China Tech Sci. 54, 3119 (2011)
15. Farquad M A H, Ravi V, Raju S B. Expert Systems with Applications. 37, 5577 (2010)
16. Baydaroğlu Ö, Koçak K. Journal of Hydrology, 508, 356 (2014)
17. Li P F, Dong L M, Xiao H C, et al. Neurocomputing. 169, 34 (2015)
18. Di Q F, Hua S, Ding W P, et al. Journal of Hydrodynamics. Ser. B. 27, 99 (2015)
19. Bai P, Zhang X.B, Zhang B. Xi'an Electronic Sience and Technology University Press (Xi'an,
2008)
20. Chang C C, Lin C J. ACM Transactions on Intelligent Systems and Technology (TIST).2, 27
(2011)
21. Chen W. Xi'an Shiyou University (Master Thesis,2014)
22. Lin W, Peng C Z, Ni X W. Well Testing, 20, 25 (2011)
23. Yang, J S. Petroleum Industry Press (Beijing, 1992)

214