Anda di halaman 1dari 11

SPE-172610-MS

Gas Lift Optimization using Artificial Neural Network


A. Ranjan, S. Verma, and Y. Singh, Indian School of MInes, Dhanbad, India

Copyright 2015, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Middle East Oil & Gas Show and Conference held in Manama, Bahrain, 8 –11 March 2015.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
Due to the increasing number of brown fields and lessening of huge oil discoveries maximizing recoveries
by artificial lift techniques can be a compensating factor to balance the demand to supply ratio. Gas lift
holds a promising aspect in this regard especially for offshore platform where every square inch costs
thousands of dollars and space is limited. Gas lift occupies very little space at the wellhead and many
directional wells can be drilled close together and easily produced. Although gas lift optimization has been
attempted using soft tools like fuzzy logic, pattern recognition and intelligent network, none of them holds
promises as big as artificial neural network. The gas lift optimization considered in this paper is to
maximize the daily hydrocarbon production by selecting optimally the lift gas rates. The procedure utilizes
well test information and calculations of vertical two phase flow behaviour to predict field producing rate
responses to changes in gas input rate. Nodal analyses, gas lift databases and gas lift monitoring system
are among the tools applied to meet the objective by training and testing various ANN models by varying
number of neurons in each layer, learning rate, training type, epoch and minimum error. It is confirmed
that in comparison with the models previously published in literature our ANN model has a better
accuracy and performance in optimizing the gas lift rate.

Introduction
Artificial lifts are used in generating flow from wells in which no flow is occurring or the flow is occurring
at an uneconomical rate. Most oil wells require artificial lift at some point in the life of the field. The major
forms of artificial lift are Sucker Rod pump, Electrical Submersible Pump (ESP), Gas Lift and intermittent
Gas Lift, Reciprocating and Jet hydraulic pumping systems, Plunger Lift and Progressive cavity pumps
(PCP). The selection of proper Artificial Lift Method is very crucial in the long run. It requires meticulous
study for proper selection of lift systems for oil fields. In every field Computer Simulation programs that
simulate the Artificial Lifts performance play an indispensible role (Khamehchi et al., 2009).
Gas lift involves injecting high-pressure gas from the surface into the producing fluid column through
one or more subsurface valves set at predetermined depths (Figure 1). This high pressure gas provides the
production well sufficient energy to flow. Gas lift can be used in deviated wellbores, and in high-
temperature environments that might have untoward effects on other lifting methods and it maximizes lift
efficiency in high-GOR wells. Further it is easy to install, cost effective and requires less maintenance.
2 SPE-172610-MS

The principle of gas lift is that gas injected into the tubing reduces the density of the fluids in the tubing,

Figure 1—Schematic diagram of a gas lift well

and the bubbles have a “scrubbing” action on the liquids. Both factors act to lower the flowing bottomhole
pressure (BHP) at the bottom of the tubing. The resultant mixture of injected gas and tubing fluid becomes
less heavy than the original oil and eventually starts to flow.
In gas lift operations, two important factors are present. First is finding of optimal position for injection
point and another is optimizing the gas injection rate (Khamehchi et al., 2009).
Our study is based on wells in India which are under gas lift operations. The following data were
garnered for the study; Input Data: Static Bottom Hole Pressure (SBHP), Flowing Tubing Head Pressure
(FTHP), Flowing Bottom Hole Pressure (FBHP), Productivity Index (PI), Tubing Size, Water Cut Percent
(WC %), Choke size, Separator temperature, Separator pressure and Gas-oil ratio; Output Data: Gas
Injection Rate (QGasInj) and Oil Production Rate (QOil).
The analytical method used in the study is based on Neural Networks, which are the latest simulation
tools in the industry. The Neural Network approach has been applied to gas lift optimization, predicting
the two output parameters of the design.
Biological Basis
Artificial neural networks have been inspired by networks of biological neuron networks. Biological
nervous system is made of interconnected processing units operating in parallel. Neurons are cells
connected to each and form a network. ANN has been inspired by human Brain which is a complex
network of biological neurons communicating with the help of each other using electrical impulses. ANNs
are generally presented as systems of interconnected “neurons” organized in different layers and neurons
of each layer are connected to the other by means of weights. These interconnected “neurons” can be
trained and used to compute values from inputs, and are capable of machine learning as well as pattern
SPE-172610-MS 3

recognition thanks to their adaptive nature(Kumar, 2012). The understanding of ANN can be made clear
by understanding the functioning of biological neuron networks.
The fundamental element of the biological neural network is a neuron. A neuron mainly consists of
three parts: dendrites, soma, and axon (Figure 2). Dendrites are the tree-like structure that receives the
signal from surrounding neurons, where each line is connected to one neuron (Figure 2). Axon is a thin
cylinder that transmits the signal from one neuron to others. At the end of axon, the contact to the dendrites
is made through a synapse. The inter-neuronal signal at the synapse is usually chemical diffusion but
sometimes electrical impulses. These impulses are made through charge carrying ions like – Na⫹, K⫹ Cl⫺
Ca2⫹. A neuron fires an electrical impulse only if certain condition is met.

Figure 2—A Biological Neuron

The incoming impulses either aid or deter firing. When the excitatory signal (aiding firing) exceeds the
inhibitory (deterring firing) signal by a certain amount in a short time period called the period of latent
stimulation, firing occurs. On assigning weight to each incoming impulse signal, the excitatory signal and
inhibitory signals have positive and negative weights respectively. This way, we can say, “A neuron fires
only if the total weight of the synapses that receive impulses in the period of latent summation exceeds
the threshold.” Figure 3 shows an example of biological neuron network in brain.
4 SPE-172610-MS

Figure 3—An example of a neural network in a human brain

Artificial Neural Network


ANN is defined as a computer model that attempts to mimic the neural network of a human brain and
simulate the specific processing of human nervous system. It is an adaptive system that establishes a
specific relation between the input and output (Singh, 2005).
Consider m number of inputs (x1, x2, x3, . . . xm) to neuron k as shown in Figure 4. The weights
connecting m number of input to kth neuron are representated by:
SPE-172610-MS 5

Figure 4 —Schematic diagram of an artificial neuron

The function of the summing junction is to collect and sum the inputs combined with the weights. The
activation function makes the neuron to produce a specific output only if it exceeds a threshold value. The
output to transfer function can be expressed as

where, bk ⫽ bias of the connection and is useful in case the summing function becomes zero.
The output of kth neuron is determined by

Network Architecture
Figure 5 depicts a network of neurons arranged in a particular fashion. The first figure is an example of
2-layer network while the second is an example of 3-layer network. Each of the connections has specific
weights and biases that is determined through a training algorithm.
6 SPE-172610-MS

Figure 5—Multilayer Neuron Network

There are two types of Artificial Neural Network: 1. Static ANN and 2 Dynamic ANN. In a static
ANN, the network is not modelled again if any error exists whereas in dynamic model, the weights and
biases are updated for better optimization using a suitable algorithm (Kumar, 2012). Dynamic model is
frequently used because of its better prediction property and has been used in this paper for the same
reason.

Conventional methodology
Gas lift optimization is an efficient method to unlock the potential of brown fields often in response to
variations in the dynamics of the reservoir and economic oscillations (Camponogara, 2005). It is an
important, but difficult and neglected task. If a well is taken with control over tubing head pressure and
increase the rate of gas in it, the production rate increases due to reduced density of the fluid in the tubing.
But as the rate is increased further, the friction drop due to the tubing becomes more important and the
production rate decreases (Figure 6). For a well with no constraints other than tubing head pressure, the
optimal rate is always the peak point. In the real world, lift gas is costly and compression cost is expressed
as a cost per unit rate of lift gas injection (for example, dollars/day per MMscf/day). This cost must be
equal to the cost of extra amount of crude oil produced. The optimum lift gas injection rate is lower than
the peak value at the point on the curve where its gradient equals the minimum economic gradient (point
B) (Khamehchi et al., 2009).
SPE-172610-MS 7

Figure 6 —Economic point and optimal point in gas lift performance curve

kGas lift optimization can be done in two ways:


1. Well by well optimization but this approach is limited by the interference from other wells and
facilities
2. Field optimization i.e. to look at the entire production systems – wells, reservoirs and surface
production systems.
In this paper, the gas lift optimization is done using the later method excluding some of the parameters
of surface network in the field with the assumption of infinite supply of lift gas.
Adopted methodology
Feed-forward backpropagation network was designed for the purpose. The bias and weights were updated
during training according to Levenberg-Marquardt optimization technique and the linear function was
used for the activation purpose.

Levenberg-Marquardt Optimization is a virtual standard in nonlinear optimization which significantly


outperforms gradient descent and conjugate gradient methods for medium sized problems.
Optimum Network Architecture
The optimum architecture for the gas lift optimization was determined on a trial and error basis. The
parameters varied were: training function, adaption learning function, transfer function, number of hidden
layers and number of neurons in each layer. The optimum number of neurons in each layer is dependent
on the complexity of the problem at hand. If the number of neurons is too few, the algorithm does not
converge to a minimum during training. Conversely, too many neurons result in over-fitting of the data
and results in poor performance. Figure 7 was selected as the optimum architecture for the gas lift
8 SPE-172610-MS

optimization. The number of hidden layer is taken as one and the number of neurons in each layer is ten.
The network takes ten input data and gives two output data as shown in the figure. The network is simple
with the minimum number of hidden layer and number of nodes as opposed to the complexity of the
prediction of gas lift optimization using conventional methods. Many other architectures were also tested
against the data but were rejected because they were unable to predict output other than the training data.

Figure 7—Optimum Network Architecture


SPE-172610-MS 9

Figure 8 —ANN best fit curve


10 SPE-172610-MS

Figure 9 —ANN regression curves (predicted outputs versus targets)

Conclusions
A single neural network was presented to predict two important parameters for the gas lift optimization.
The first parameter determines the injection rate of the gas at which the rate of oil production is maximum
and the second parameter determines this maximum rate of oil production. The training data used in the
literature was taken from the entire field and is better than well by well optimization technique. Moreover,
the transfer function used in our literature is a simple linear function and hence, indicates a linear
relationship between the input and output parameters. Lastly, the architecture of the network is simple
with only 1 hidden layer and 10 neurons in the hidden layer. In a nutshell, the simplicity of the network
makes it faster to compute than the models previously published in the literature.
It does not eliminate the need for the conventional technique but automates the process. It saves the
engineer from tedious calculations of the nodal analysis and hence speeds up the calculations.
SPE-172610-MS 11

Over the years, the oil prices have varied erratically and the company is forced to utilize the existing
oil resources of the field. Gas Lift Optimization using Artificial Neural Network makes a significant
contribution towards this direction.

Acknowledgements
The authors would like to thank Dept. of Petroleum Engineering, Indian School of Mines, Dhanbad for
useful discussion and cooperation on this research. The authors are indebted to Prof. V.P. Sharma,
Professor and Head of Department, Dept. of Petroleum Engineering, Indian School of Mines, Dhanbad,
Dr. Keka Ojha, Associate Professor, Dept. of Petroleum Engineering, Indian School of Mines, Dhanbad,
Dr. Vikas Mahto, Associate Professor, Dept. of Petroleum Engineering, Indian School of Mines, Dhanbad,
and officials from Schlumberger for their noble guidance. The authors extend their thanks to Mr. Vinay
Kumar Rajak, Assistant Professor, Dept. of Petroleum Engineering, Indian School of Mines, Dhanbad
whose encouragement and support has always been inspiring and rejuvenating.

References
1. Babu, D.K. and Odeh, A.S. 1989. Productivity of a Horizontal Well. SPE Res Eng 4(4): 417–421.
SPE-18298-PA.
2. Beale, M.H. Hagan, M.T. and Demuth, H.B. 2013. Neural Network ToolboxTM User’s Guide,
The MathWorks Inc., MA, USA.
3. Bennett, P. 1980. Artificial Lift Concepts and Timing. Petroleum Engineer (May): 144.
4. Chia, Y.C. and Hussain S., 1999. Gas Lift Optimization Efforts and Challenges, SPE 57313, SPE
Asia Pacific Improved Oil Recovery Conference, Kaula Lumpur, Malaysia, 25-26 October.
5. Camponogara, E., 2005. Solving a gas-lift optimization problem by dynamic programming. Eur.
J. Operat. Res., 174: 1220 –1246.
6. Khamehchi, E., Rashidi, F. and Rasouli, H. 2009. Prediction of Gas lift parameters using Artificial
Neural Network. Enhanced Oil Recovery – Iranian Chemical Engineering Journal (Special Issue)
– Vol. 8 – No. 43.
7. Khamehchi, E., Rashidi, F., Omranpour, H., Ghidary, S.S, Ebrahimian, A. and Rasouli, H. 2009.
Intelligent System for Continuous Gas Lift Operation and Design with Unlimited Gas Supply.
Journal of Applied Sciences, 9: 1889 –1897.
8. Kumar, A. 2012. Artificial Neural Network as a Tool for Reservoir Characterization and its
Application in the Petroleum Engineering, OTC 22967, Offshore Technology Conference, Hous-
ton, 30 April – 3 May.
9. Langston, L.V., 2003. The Lease Pumper’s Handbook, First Edition. Commission on Marginally
Producing Oil and Gas Wells, State of Oklahoma.
10. McCulloch, W.S. and Pitts, W. 1943. A Logical Calculus of the Ideas Immanent in Nervous
Activity, Bulletin of Mathematical Biophysics, 5.
11. Redden, J.D., Sherman, T.A.G and Blann, J.R., 1974. Optimizing Gas Lift Systems, SPE 5150,
49th Annual Fall Meeting of the Society of Petroleum Engineers of AIME, Houston, Texas, 6-9
October.
12. Singh, S. 2005. Permeability Prediction using Artificial Neural Network (ANN): A Case Study of
Uinta Basin, SPE – 99286-STU, SPE ATCE, Dallas, 9 – 10 October.
13. Takacs, G. 2010. Ways to Obtain Optimum Power Efficiency of Artificial Lift Installations. SPE
126544, Oil and Gas India Conference and Exhibition, Mumbai, India, 20 –22 January.
14. Takacs, G. 2005. Gas Lift Manual. PennWell Books, Tulsa, Oklahoma.
15. Vogel, J.V. 1968. Inflow Performance Relationships for Solution-Gas Drive Wells. Journal of
Petroleum Technology 20(1): 83–92. SPE 1476-PA.

Anda mungkin juga menyukai