Group Title: 7th International Conference on Multiphase Flow - ICMF 2010 Proceedings
Title: P1.43 - Pressure Drop Prediction of Oil-gas-water in Horizontal Pipeline with Super-high Water-cut Based on BP Neural Network
ALL VOLUMES CITATION THUMBNAILS PAGE IMAGE ZOOMABLE
Full Citation
STANDARD VIEW MARC VIEW
Permanent Link: http://ufdc.ufl.edu/UF00102023/00451
 Material Information
Title: P1.43 - Pressure Drop Prediction of Oil-gas-water in Horizontal Pipeline with Super-high Water-cut Based on BP Neural Network Industrial Applications
Series Title: 7th International Conference on Multiphase Flow - ICMF 2010 Proceedings
Physical Description: Conference Papers
Creator: Liu, X.
Liu, J.
Liu, L.
Mao, Q.
Publisher: International Conference on Multiphase Flow (ICMF)
Publication Date: June 4, 2010
 Subjects
Subject: three-phase flow
super high water-cut
pressure drop prediction
neural network
MATLAB
 Notes
Abstract: With persistent exploitation of oilfield, the water-cut of oil wells output rises continuously. The research of pressure drop of oil-gas-water mixed transportation pipeline is very significant for the operation and management of oil-gas gathering and transferring system. In this paper, the pressure drop is tested with different pipe lengths, pipe diameters, liquid flow, gas flow, water-cut and temperatures. Basing on the test data, with MATLAB toolbox, the BP neural network program is developed to predict pressure drop in oil-gas-water horizontal pipeline with super-high water-cut. The predicting results indicate that the BP neural network method given in this paper is feasible.
General Note: The International Conference on Multiphase Flow (ICMF) first was held in Tsukuba, Japan in 1991 and the second ICMF took place in Kyoto, Japan in 1995. During this conference, it was decided to establish an International Governing Board which oversees the major aspects of the conference and makes decisions about future conference locations. Due to the great importance of the field, it was furthermore decided to hold the conference every three years successively in Asia including Australia, Europe including Africa, Russia and the Near East and America. Hence, ICMF 1998 was held in Lyon, France, ICMF 2001 in New Orleans, USA, ICMF 2004 in Yokohama, Japan, and ICMF 2007 in Leipzig, Germany. ICMF-2010 is devoted to all aspects of Multiphase Flow. Researchers from all over the world gathered in order to introduce their recent advances in the field and thereby promote the exchange of new ideas, results and techniques. The conference is a key event in Multiphase Flow and supports the advancement of science in this very important field. The major research topics relevant for the conference are as follows: Bio-Fluid Dynamics; Boiling; Bubbly Flows; Cavitation; Colloidal and Suspension Dynamics; Collision, Agglomeration and Breakup; Computational Techniques for Multiphase Flows; Droplet Flows; Environmental and Geophysical Flows; Experimental Methods for Multiphase Flows; Fluidized and Circulating Fluidized Beds; Fluid Structure Interactions; Granular Media; Industrial Applications; Instabilities; Interfacial Flows; Micro and Nano-Scale Multiphase Flows; Microgravity in Two-Phase Flow; Multiphase Flows with Heat and Mass Transfer; Non-Newtonian Multiphase Flows; Particle-Laden Flows; Particle, Bubble and Drop Dynamics; Reactive Multiphase Flows
 Record Information
Bibliographic ID: UF00102023
Volume ID: VID00451
Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
Resource Identifier: P143-Liu-ICMF2010.pdf

Full Text


7th International Conference on Multiphase Flow
ICMF 2010, Tampa, FL USA, May 30-June 4, 2010




Pressure Drop Prediction of Oil-gas-water in Horizontal Pipeline with Super-high
Water-cut Based on BP Neural Network


Xiaoyan Liu, Jiajia Liu, Lijun Liu and Qianjun Mao

Daqing Petroleum Institute, Civil and Architecture Engineering, Department of Built Environment, Thermal Power
Engineering
199, FaZhan Street, SaErTu, Daqing, 163318, China
renenegoncheng@1~26. com


Keywords: Three-phase flow: Super high water-cut, Pressure drop prediction, Neural Network, MATLAB




Abstract

With persistent exploitation of oilfield, the water-cut of oil wells output rises continuously. The research of pressure drop of
oil-gas-water mixed transportation pipeline is very significant for the operation and management of oil-gas gathering and
transferring system. In this paper, the pressure drop is tested with different pipe lengths, pipe diameters, liquid flow: gas flow:
water-cut and temperatures. Basing on the test data, with MATLAB toolbox, the BP neural network program is developed to
predict pressure drop in oil-gas-water horizontal pipeline with super-high water-cut. The predicting results indicate that the BP
neural network method given in this paper is feasible.


Introduction

With the deep exploitation of oil field, the eastern oil fields
have entered high water-cut period [1]. The water-cut has
exceeded 85% in many oil fields. Predicting pressure drop
exactly could provide a reliable basis for oil field production
and design of gathering-transporting pipeline.
After long-range research, people have developed some
methods to predict the three-phase flow pressure drop,
including two aspects: some people developed three-phase
flow pressure drop theory on the basis of accurately
analyzing oil-water mixture physical properties and
applying mechanically gas-liquid two-phase flow pressure
drop theory [2-5]; others established two-fluid or three-fluid
hydrodynamic models and analyzed three-phase flow
pressure drop from a view of fluid mechanics [6-10]. Liu
Xiaoyan from Daqing Petroleum Institute set up a suit of
equipment to study the questions of pipeline pressure drop
with produced fluid in super-high water-cut stage and then
corrected Baker model with testing data. The testing results
indicated that the modified model was fit for calculating the
slug flow pressure drop of oil-gas-water three-phase
horizontal pipeline with super high water-cut.
In this paper, we will test the pressure drop under different
conditions using the on-site testing equipment developed by


research group. Basing on the test data, a BP neural network
program for predicting pressure drop in oil-gas-water
horizontal pipeline with super-high water-cut is going to be
given.

Acquisition of the sample data

The test equipment is installed in Daqing oilfield, and the
process chart of the device is shown in Fig.1.
Taking 617# well hot watered as example, the measuring
procedure is illustrated. (1) When the gas or liquid
production is tested, the valve 1# is opened. The production
fluid of 617# well is separated into oil and gas in separating
tank, and the gas production rate and liquid producing
capacity are measured. Close the valve 4#, open the valve
3#, 5#, 6#, and test the water that injected to 617#well, other
valves keep normal operation. (2) When test the flow
pattern and flow state, valve 1# and 8# are opened, the valve
10# for injecting is closed, valve 3#, 5#, 6#, 7# are all
closed, valve 9# is closed. The oil-gas-water mixture flows
into the flow pattern testing equipment, and passes through
the pressure gauge, temperature gauge, then get into the
header by passing valve 9#.





















































well number 217# 313# 614# 617# 316# 616# 713#
length/ 70 319 259 90 186 175 454
diameter/mm 76 76 60 60 89 60 60
dphmm 110 500 500 560 500 300 700
Liquid ouptm.d ~ 84.61 34.96 25.94 25.06 27.19 68.70 40.54
Gas ouptm .d ~ 1457.86 134.12 397.90 100.19 142.52 358.15 188.36
water-cut/% 92.7 92.3 91.2 87.9 78.0 90.0 90.8
Gas oil ratio/m /t 236.03 49.82 174.31 33.04 22.83 52.13 50.50


Table 2: The test data of mutpaeflow in Gathering Ppline of 617#well
Liquid flux (m .d ') Gas flux (m .d Water-cut (%) Aergtmpaue (oC) Pressure drp(MPa)
53.83 79.55 92.3 30.29 0.090
23.30 75.75 86.1 28.14 0.078
30.63 75.41 87.1 27.77 0.047
25.02 95.51 90.7 27.82 0.078
20.05 95.51 87.4 27.79 0.087
23.00 84.35 90.7 27.37 0.099
23.12 73.13 87.0 27.41 0.076
25.25 114.8 86.6 27.74 0.076
32.27 136.94 91.0 36.29 0.067
40.14 140.97 92.4 29.43 0.079
35.26 130.22 91.5 30.05 0.055


7th International Conference on Multiphase Flow
ICMF 2010, Tampa, FL USA, May 30-June 4, 2010


Middle Station


3


Middle Station


{ J2Fhree 11ays 1 alvel


-Gas Pipe


Sah e


Llq~uid Pressure Temperature Heat 11ater
Measure Meter Measure Pipe


-Oi1l Ppe Testng
Pipe


Figure 1: The technical flow chart of the test device in Daqing Oilfield


Test program: test the temperature and pressure of single
well collecting line well head and the oil-gas-water mixture
back to metering plant respectively. And liquid producing
capacity, gas production rate and water cut of the well are
measured by gauging and sampling. The length and
diameter of pipeline between the well head and metering

Table 1: The sum-up table of testing wells' basic circumstances


plant are re-measured. The average value of flowline
temperature and scavenge pipe temperature is the average
temperature of the pipeline, and the difference of flowline
pressure and scavenge oil pressure is the pressure drop of
pipeline. The situations of test wells and relevant gathering
lines are shown as tablel.


Take the corresponding pipeline characteristics of 617#well


with high water-cut as example, as shown in table 2.






7th International Conference on Multiphase Flow
ICMF 2010, Tampa, FL USA, May 30-June 4, 2010

Design of hidden layer

There is an important theorem for the BP network [11], that
is, any continuous function within a closed interval can be
approximated with a single layer BP network, thus a
three-layer network can accomplish any mapping from
n-dimension to m-dimensional. For the prediction of
preSsure drop of oil-gas-water transmission pipes,
three-layer BP neural network structure can be adopted, that
is the input layer-a single hidden layer-output layer, in
which the units number of input layer is 6, the units number
of output layer is 1. It is a very complex issue to determine
neurons number of hidden layers, it has some straight
relationship with question demand and the nerve cell
number of input and output, which is usually based on
designers experience and numerous experiments [12]. The
referenced formula to select the best number of neurons is
shown as (2)
n, = ~+ a (2)
Where: n; is the neurons number of hidden layer; n is the
11CUTORS number of input layer 6; m is the neurons number
of output layer 1; a is a constant between 1 to 10, then the
number of hidden layer neurons is between 3~13. After
repeated training (the activation function of hidden layer is
logsig, the activation function of hidden layer is purelin)
and comparison the situations of network approximation in
different number of neurons, the number of hidden layer
neurons is fixed 13.

Select Training Methods

During the process of BP network training, it often occurs
that new input will make the network training error increase
rapidly when the network training error is very small, which
is due to the network without good generalizing ability, that
is, there is large error between the output for new input and
the corresponding target output of the network. One way to
improve the network generalizing ability is increasing the
size of the network properly. However, for a specific
problem, it is very difficult to define the size of the network
well beforehand. Wherefore, we enhance the generalizing
capacity of network by adjusting the performance function
of network [13]. The performance function for ordinary BP
neural network is MSE, such as the type (3) as follows:
1 1 (3)
mse n(e,)2= Cx I ~2

Where: e,, t,, a, are the training error, the target output and
network output of the sample i respectively. The adjusted
function of network performance is showed as type (4):
msereg = mnse +(1- p)msw (4)

1 2 )' 5)

Where: pu is the performance parameters, w, is the weight of
network. Using pu can reduce the weight and the threshold of
the network effectively to make the training output of the
network become smoother, which can increase the network
generalization performance. But if the performance
parameters pu is too large, the network generalization ability
would be bad; if it is too small, the training accuracy is too
low. MATLAB neural network toolbox provides function
trainbr which could set the optimal performance parameters


Design of BP Neural Network

Neural network, whose full name is artificial neural
network, is a kind of artificial intelligence algorithm
which uses large number of simple calculation elements
to simulate a biological neural system. The structure of
BP neural network is one-way multi-layer, as shown in
Fig.2. It generally consists of an input layer (py, p2, *** ,
pn), one or more hidden layers and one output layer (ty,
t2, ..., t,). Each layer contains a number of neurons, the
neurons at the same level are independent on each other,
and the information among layers is transmitted along
one direction. The core idea of BP algorithm is adjusting
the network weight by output layers error feedback
hidden layers error, which makes the output signal attain
some accuracy.

pi t' 1 (1 L t


p2 t2 v l




pn I to

Figure 2: The structure of BP neural network

Select sample point

The test data which have influence on the pressure drop of
oil-gas-water in pipeline, including tube length, diameter,
liquid flux, water-cut, gas flux and the temperature, are
selected as six-dimensional input parameters of the network,
the pressure drop of three-phase flow under the
corresponding condition is concerned as the output
parameter.

Data Normalization

The physical quantities of BP network input nodes are
different, and some values vary greatly. In order to prevent
the large numerical from inundating small value of
information, we can normalize the input data from 0 to 1.
The change of sigmoid function value in BP algorithm is
extremely slow when the input parameter is close to 0 and 1,
which easily causes the common phenomenon of network
paralysis. In order to reduce learning time, we control the
data of input and output at the range from 0.1 to 0.9, so that
the variable gradient of sigmoid function value is relatively
large at this range, the network convergence time is
shortened greatly with improving network performance, the
transforming method is shown as (1).
x' = 0.1+ 0.8 x x xmm (1)
xma xmm
Where: x is the data before normalized; xmax, xman are the
maximum and the minimum values of the data before
normalized; x 'is normalized data respectively.









automatically, so the training function in tlus paper is
trainbr.

Network training and testing

Take the test data in Daqing Oil Field as samples, a total of
95 groups, 75 samples are used to train by the designed BP
network, the MSEREG reaches 0.001 through 361 times
training, at last the learning and training is finished.
Comparing the simulating result with testing data, the
curves are shown in Fig. 3.

-* Tested vahw
0 3 -- Sirnulatedralue



a 1



1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73
Trained samples
Figure 3: The pressure drop comparison between trained
results and test results of samples

In the command window of MATLAB software
environment, makes prediction by the workout BP network
with the remaining 20 groups samples, the results are shown
in Fig. 4.


0.3 -*-Testedvalue
0.25 =Predictedvalue

S0.2 -
80.15 -


0.05 -


1 5 7 ~~Pred~icted slampl~es15 719

Figure 4: The pressure drop comparison between
predicted results and test results of samples

Error analysis is shown in Table 3.
Table 3: The error contrast and analysis between predicted
results and test results
Error scope Number ofgrus
10% 7
10%~i~20% 9
> 20% 4

As seen in Table 3, the generalization ability and
convergence speed of designed BP neural network in this
article is good. The results show that errors within 10%
occupy 35%, and errors within 100Ard-t20% occupy 45%.
In a word, since the relative error of 80% within & 20%, we
conclude that the BP neural network given in this paper is fit
for the pressure drop predicting of oil-gas-water three-phase
horizontal pipeline with super high water-cut.


7th International Conference on Multiphase Flow
ICMF 2010, Tampa, FL USA, May 30-June 4, 2010


Conclusion

(1)By using of the devices designed by our project group,
the pressure drop is tested for the pipelines in different
conditions of pipe length, pipe diameter, fluid flow: gas flow:
water-cut and temperature.

(2) Based upon the MATLAB Toolboxes, a BP neural
network program is brought for predicting pressure drop in
oil-gas-water horizontal pipeline with super-high water-cut.
A three-layer BP neural network structure is adopted in the
program. The test data which influence the pressure drop of
oil-gas-water in pipeline are selected as six-dimensional
input parameter of the network, which include pipe length,
pipe diameter, liquid flow: gas flow: water-cut and
temperature: the pressure drop of three-phase flow is
selected as the output parameter. The activation function of
hidden layer is logsig: the activation function of hidden
layer is purelin: the training function is trainbr.

(3) The network testing results with samples indicate that
for 80% the data, the relative error of is less than 20%.
Therefore, the BP neural network is fit for the slug flow
pressure drop prediction of oil-gas-water three-phase
horizontal pipeline with super high water-cut.

Acknowledgements

Due thanks are given to Xuying for the help during the
process of writing.

References

LIU Xiao-Yan, ZHANG Yan, ZHAO Bo, et al. Test of
Pressure Drop in Gathering and Transfering Pipelines.
Oil-Gas field Surface Engineering, Vol. 24 (2005)

Malinowsky M. An Experimental Study of Oil-Water and
Air-Oil-Water Flowing Mixtures in Horizontal Pipes, M S
Thesis, The U. of Tulsa, 1975

Stapelberg H, Dorstewita F, Nadler M. Slug Flow of Oil,
Water and Gas in Horizontal Pipelines. Elsevier Science
Publishers, 1991

Aziz K et al. Intermittent Two Phase Flow Predictive
Models in Horizontal Pipes. Can. J. Chem. Eng, Vol. 56
(1978)

Hall A RW et al. An Experimental Investigation of the
Water Phase in the Multiphase Flow of Oil, Water and Gas.
Proc. 6th Int. Conf. on Mutilphase Production, France, 1993

Neogi S, Lee A, J epson W P. A Model for Multiphase
(gas-water-oil) Stratified Flow in Horizontal Pipeline, SPE
28-99 (1994)

Taitel Y, Bamnea D, Brill J P. Stratified Three Phase Flow in
Pipes. Int. J ofMultiphase Flow, Vol. 21 (1995)

Hewitt G F, Hall A, Pan L. Three-Phase Gas-Liquid-Liquid
flow, A New Challenge. First Intemnational Symposium on






7th International Conference on Multiphase Flow
ICMF 2010, Tampa, FL USA, May 30-June 4, 2010

Two-phase Flow Modeling and Experimentation, Italy,
1995

Hewitt G F, Pan L, Khor A H: Three-Phase
Gas-Liquid-Liquid Flow: Flow Pattern, Holdup and
Pressure Drop, ISMF, BeiJing, 1997

Bonizzi M, Issa R I. On the Simulation of Three-Phase Slug
Flow in Nearly Horizontal Pipes Using the Multi-Fluid
Model, Int. J. of Multiphase Flow, Vol. 29 (2003)

Hagan T, Demuth H, Beale M. Neural Network Design.
Boston, MA, PWS Pulishing, 1996

FEISI Technology Product Research Center. Neural
network theory and implementation based upon MATLAB.
Electronics Industry Press, Vol. 3, 99-107 (2005)

SHANG Jie, ZHU Zhan-Li. Improving of learning speed
and generalization capability of BP neural network. Fujian
computer, Vol. 4, 20-21 (2005)




University of Florida Home Page
© 2004 - 2010 University of Florida George A. Smathers Libraries.
All rights reserved.

Acceptable Use, Copyright, and Disclaimer Statement
Last updated October 10, 2010 - Version 2.9.7 - mvs