SPE 133436

Use of Fuzzy Logic for Predicting Two-Phase Inflow Performance Relationship

of Horizontal Oil Wells
M. Ebrahimi, SPE, ACECR-Production Technology Research Institute, and A. Sajedian, SPE, Kish Petroleum

Copyright 2010, Society of Petroleum Engineers

This paper was prepared for presentation at the Trinidad and Tobago Energy Resources Conference held in Port of Spain, Trinidad, 27–30 June 2010.

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
A straight line equation is generally used to estimate well inflow performance above bubble-point pressure. However, when the
pressure drops below the bubble-point pressure, the trend deviates from that of the simple straight line relationship. Although some
analytical methods can accurately represent the horizontal well IPR behavior above bubble point pressure, only empirical
correlations are available for IPR modeling of two-phase reservoirs and hence some deviations from actual data are often
observed.
Artificial intelligence techniques such as neural networks, fuzzy logic, and genetic algorithms are increasingly powerful and
reliable tools for petroleum engineers to analyze and interpret different areas of oil and gas industry. In this paper, two neuro-fuzzy
models, including Local Linear Neuro-Fuzzy Model (LLNFM) and Adaptive Neuro Fuzzy Inference System (ANFIS) have been
compared with Multi-Layer Perceptron (MLP) and empirical correlations to predict the inflow performance of horizontal oil wells
experiencing two phase flow.
Several reservoir models have been simulated with different bottomhole pressures. The models contained a wide range of
absolute and relative permeabilities, PVT data, and horizontal well lengths. The necessary training data have been obtained from
80% of simulation results, covering a wide range of fluid and rock properties. The other 20% are used for error checking and
performance testing. The results show that the Local Linear Neuro-Fuzzy Model gives the smallest error for unseen data, when
compared to other intelligent models and empirical correlations.

Introduction
A commonly used measure of the ability of the well to produce is the Productivity Index. Defined by the symbol J, the
productivity index is the ratio of the total liquid flow rate to the pressure drawdown. For a water-free oil production, the
productivity index is given by (Ahmed, 2001):

Qo
J= (1)
P r − Pwf

Where J is the productivity index and Pr is the average drainage area pressure.
Equation (1) can alternatively be expressed as (Ahmed, 2001):

⎛1⎞
Pwf = P r − ⎜⎜ ⎟⎟Qo (2)
J ⎝ ⎠
2 SPE 133436

The above expression which shows the plot of Pwf against Qo is a straight line with a slope of (-1/J) as shown schematically in
Figure 1. This graphical representation of the relationship that exists between the oil flow rate and bottomhole flowing pressure is
called the inflow performance relationship and referred to as IPR (Ahmed, 2001).

Figure 1: Inflow performance curve

IPR is used for evaluating reservoir deliverability in production engineering. Reservoir deliverability determines types of
completion and artificial lift methods to be used. Reservoir deliverability can be mathematically modeled on the basis of flow
regimes such as transient flow, steady state flow, and pseudo–steady state flow (Gue et al., 2007).
The linear IPR model is valid for pressure values as low as bubble-point pressure. Below the bubble-point pressure, the
solution gas escapes from the oil and become free gas. The free gas occupies some portion of pore space, which reduces flow of
oil. This effect is quantified by the reduced relative permeability. Also, oil viscosity increases as its solution gas content drops. The
combination of the relative permeability effect and the viscosity effect results in lower oil production rate at a given bottomhole
pressure (Gue et al., 2007).
Only empirical equations are available for IPR modeling of two-phase reservoirs. In this paper, the ability of the artificial
intelligence in establishing and predicting well inflow performance for solution gas drive reservoirs producing through a horizontal
well is to be investigated. A wide range of data is needed to build artificial intelligence models. 18 different numerical models for a
horizontal solution gas drive reservoir with variable PVT, relative permeability, and well length characteristics were built and
simulated, assuming different constant bottomhole pressures. The results are then used to build three artificial intelligence models
that can generate a predictive IPR curve.

Literature Review
For vertical single phase oil wells, Darcy’s equation can easily be used to evaluate IPR for steady and pseudo-steady state flows.
For single phase horizontal oil wells, the simplest form of well productivity calculations was presented by Joshi (1991) for steady-
state flow. Many authors such as Borisov (1964), Giger et al. (1984), and Renard and Dupuy (1990) have presented some
equations for pseudo-steady state flow, which are similar or based on the Joshi’s equation for steady state flow. However, only
empirical equations are available for IPR modeling of two-phase reservoirs.
Vogel’s equation is the simplest and the most widely used IPR prediction equation in the petroleum industry. Vogel (1968)
used a numerical simulation to investigate the IPR behavior of a solution-gas-drive reservoir producing below the bubble point
pressure. He generated IPR curves for 21 fictitious solution-gas-drive reservoirs that covered different fluid PVT properties and
relative permeability characteristics. Vogel plotted the different IPR curves as dimensionless IPR curves, where for a particular
curve the pressure was divided by the maximum or shut-in pressure, and the corresponding flow rate was divided by the maximum
flow rate. When he noticed that the curves were generally exhibiting a similar shape, he applied regression analysis to find the best
curve that can represent all curves. A dimensionless IPR curve could be used to create the IPR curve for any well that produces
from a solution-gas-drive reservoir at any depletion stage, using only a one-point test. The equation for this dimensionless IPR
curve is (Vogel, 1968):
SPE 133436 3

2
Qo Pwf ⎛ Pwf ⎞
= 1 − 0.2 − 0.8⎜ ⎟ (3)
(Qo ) max Pr ⎜ P ⎟
⎝ r ⎠

Fetkovich (1973) found that the back-pressure curves for the data collected from saturated and undersaturated reservoirs follow
the same general form as that used for gas wells as follows:

n
⎛ 2 2⎞
Q = C ⎜ Pr − Pwf ⎟ (4)
⎝ ⎠

Where n is the exponent of back-pressure curve.

The complex flow regime existing around a horizontal wellbore probably precludes using a method as simple as that of Vogel
to construct the IPR of a horizontal well in solution gas drive reservoirs. If at least two stabilized flow tests are available, however,
the parameters J and n in the Fetkovich equation (Equation 4) could be determined and used to construct the IPR of the horizontal
well. In this case, the values of J and n would not only account for effects of turbulence and gas saturation around the wellbore, but
also for the effects of non-radial flow regime existing in the reservoir (Ahmed, 2001).
Bendakhlia and Aziz (1989) used a reservoir model to generate IPRs for a number of wells and found that a combination of
Vogel and Fetkovich equations would fit the generated data if expressed as:

n
⎡ 2⎤
Qoh ⎢ ⎛ Pwf ⎞ ⎛P ⎞ ⎥
= ⎢1 − V ⎜ ⎟ − (1 − V )⎜ wf ⎟ (5)
(Qoh )max ⎜ Pr ⎟ ⎜ P ⎟ ⎥
⎢ ⎝ ⎠ ⎝ r ⎠ ⎥
⎣ ⎦

Where (Qoh) max = horizontal well maximum flow rate, STB/day

n = exponent in Fetkovich’s equation
V = variable parameter

In order to apply the equation, at least three stabilized flow tests are required to evaluate the three unknowns (Qoh)max, V, and n
at any given average reservoir pressure. However, Bendakhlia and Aziz (1989) indicated that the parameters V and n are functions
of the reservoir pressure or recovery factor and, thus, the use of Equation 5 is not convenient in a predictive mode.
Cheng (1990) presented a form of Vogel’s equation for horizontal wells that is based on the results from a numerical simulator.
The proposed expression has the following form:

2
Qoh ⎛ Pwf ⎞ ⎛P ⎞
= 1.0 + 0.2055⎜⎜ ⎟ − 1.1818⎜ wf ⎟ (6)
(Qoh )max ⎟ ⎜ P ⎟
⎝ Pr ⎠ ⎝ r ⎠

Retnanto and Economides (1998) proposed the following correlation:

n
Qoh ⎛ Pwf ⎞ ⎛P ⎞
= 1.0 − 0.25⎜⎜ ⎟ − 0.85⎜ wf ⎟ (7)
(Qoh )max ⎟ ⎜ P ⎟
⎝ Pr ⎠ ⎝ r ⎠
4 SPE 133436

⎛ 2⎞
⎜ ⎛P ⎞ ⎛P ⎞ ⎟ −3 ⎞
n = ⎜ − 0.27 + 1.46⎜ r ⎟
− 0.96⎜ r ⎟ ⎛
⎟⎜ 4.0 + 1.66 × 10 Pb ⎟ (8)
⎜ ⎜ P ⎟ ⎜ P ⎟ ⎝ ⎠
⎝ b⎠ ⎝ b⎠ ⎟
⎝ ⎠

Procedure
The ability of the artificial intelligence in establishing and predicting well inflow performance for solution-gas-drive reservoirs
with a horizontal well is to be investigated. A wide range of data is needed to build artificial intelligence models. Eighteen
different numerical models for a solution gas drive reservoir with variable PVT properties, relative permeability characteristics,
and well lengths were built and simulated, assuming different constant bottomhole pressures. The results are then used to build
three artificial intelligence models that can generate a predictive IPR curve. In the learning process, the models used both input and
output data, taken from reservoir simulations, to construct an artificial intelligence model. The model then predicted the output
only from the input data.

Model Description
The simulated model is a box-shaped solution-gas-drive reservoir with 2000 ft long, 1400 ft wide, and 300 ft deep. There are
20 cells in the x-direction, 11 cells in the y-direction, and 9 cells in the z-direction. Since fluid flow and pressure profile change
more rapidly near wellbore, cell sizes are arranged in a way that smaller cell sizes lie near the wellbore. The reservoir is
anisotropic and the permeability in z-direction is one-tenth of the permeability in x and y directions. There is a single producing
horizontal in the center of the grid. The basic reservoir properties are given in table 1 and the following assumptions were
considered for simplification:

1) The oil zone is one layer and has constant thickness for the whole reservoir.
2) The reservoir is completely bounded.
3) The initial water saturation is at the connate water saturation.
4) Gravity and capillary pressure effects were neglected.
5) Initial reservoir pressure is taken to be equal to bubble point pressure

Table 1: Reservoir properties

Porosity (φ) 25%
Horizontal permeability (Kh) 16 md
Formation thickness (h) 300 ft
Well radius (rw) 0.51 ft
Outer radius (re) 1000 ft

Three different relative permeability curves (figure 2), three different fluid properties (figure 3), and two different well lengths
(400 ft, and 1400 ft) were used to cover a wide range of cases. These properties made 18 distinct simulation models; each
simulated assuming different constant bottomhole pressures. This provided us with the necessary input parameters for our artificial
intelligence models.

st st
Figure 2(a): Gas relative permeability- 1 case Figure 2(b): Oil relative permeability- 1 case
SPE 133436 5

nd nd
Figure 2(c): Gas relative permeability- 2 case Figure 2(d): Oil relative permeability- 2 case

rd rd
Figure 2(e): Gas relative permeability- 3 case Figure 2(f): Oil relative permeability- 3 case

Figure 3(a): Oil viscosity against pressure

Figure 3(c): Oil formation volume factor against pressure

6 SPE 133436

Figure 3(b): Inverse of gas formation volume factor vs. pressure

Figure 3(d): Gas solubility against pressure

Simulation and Data Preparation

Each of the 18 distinct simulation models was run with different constant bottomhole pressures. Of the results, oil flow rate
(Qo) and its corresponding cumulative oil produced (Np) versus time was of interest. The data for 5500 time steps were extracted
and arranged to build the artificial intelligence models.
For the learning process of the neural network, both input and output parameters were involved. An IPR curve requires only the
values of the bottomhole pressure and its corresponding oil flow rate, but defining a specific IPR curve requires more parameters.
These input parameters should have a relationship with the output, according to which, we used the following simulation input and
output parameters as the input layer of artificial intelligence models:

a) The recovery factor, Np / N

b) Flowing bottomhole pressure, Pwf
c) Average reservoir pressure, Pr
d) Maximum oil flow rate, Qo, max
e) Bubble point pressure, Pb
f) Oil formation volume factor at bubble point pressure, Bob
g) Solution gas-oil ratio at bubble point pressure, Rsb
h) Gas viscosity at bubble point pressure, μgb
i) Well length

We used the oil and gas properties at bubble point pressure because it is difficult to involve those properties at every pressure,
as they are changing. Properties at the bubble point, however, would be fixed for the whole life of the reservoir. The output for
artificial intelligence models is selected to be the oil flow rate, Qo. Three different types of artificial models were designed and
their accuracy against actual simulation data was examined.
SPE 133436 7

Results and Discussion

First Model (Multi-Layer Perceptron)

For the first model it is tried to build a neural network model, of Multi-layer Perceptron (MLP) type that could predict the oil
flow rate by including all the effective parameters, whether directly or indirectly. 14 out of 18 simulation models were used for
network training and the other two models were used for testing and validating the constructed network. Figures 4 and 5 show the
training and testing process for this model. Different numbers of neurons were tested in hidden layers. The best combination was
selected to be the one with two hidden layers, each of three and five neurons. This model, however, can not accurately predict IPR
curves and has a large deviation from actual data as much as 3.5%.

Figure 4: Training process for MLP model

Figure 5: Testing process for MLP model

Second Model (ANFIS)

In the second model we tried to build a Fuzzy neural network model called Adaptive Neuro Fuzzy Interface (ANFIS) that uses
fuzzy clustering method for extracting rules. This model could predict IPR curves better than the previous model with a deviation
of 1.2% from actual data. Figures 6 and 7 show the training and testing processes for this model.
8 SPE 133436

Figure 6: Training process for ANFIS model

Figure 7: Testing process for ANFIS model

Third Model (LLNFM)
For the third model, another kind of neuro-fuzzy network called Local Linear Neuro-Fuzzy Model (LLNFM) is built that uses a
new enhanced neuro-fuzzy training approach called local linear model tree (LOLIMOT) algorithm for network training. Figures 8
and 9 show the training and testing processes for this model. The model could predict IPR curve accurately with an error of only
0.05%, nominating this model as the most accurate one. Figure 10 compares different IPR models in terms of mean square error.

Figure 8: Training process for LLNFM model

SPE 133436 9

Figure 9: Testing process for LLNFM model

Figure 10: Mean square error for different IPR models

Conclusions
Although IPR can be analytically expressed by the Darcy’s law above bubble point pressure, only empirical expressions are
available to estimate IPR below bubble point. Two neuro-fuzzy models, Local Linear Neuro-Fuzzy Model (LLNFM) and Adaptive
Neuro Fuzzy Inference System (ANFIS), have been compared with Multi-Layer Perceptron (MLP) and empirical correlations to
predict the inflow performance of horizontal oil wells experiencing two phase flow. Eighteen different reservoir models with a
wide range of relative permeability, PVT properties, and horizontal well lengths have been run in different bottomhole pressures.
The data obtained from 14 models have been used for training. The other four models are used for error checking and performance
testing. The results show that the Local Linear Neuro-Fuzzy Model based on LOLIMOT algorithm gives the smallest error for
unseen data, when compared to other intelligent models and empirical correlations. This method is preferred because not only it is
accurate but also is fast and forward (no trial and error needed).

Nomenclatures

Φ= porosity
h= formation thickness, ft
J= Productivity index, STB/day/psi
n= exponent of back pressure curve
Pb = bubble point pressure, psi
Pr = Average drainage area pressure, psi
Pwf = Bottomhole flowing pressure, psi
Qo = Oil flowrate, STB/day
Qoh = Horizontal well oil flowrate, STB/day
(Qoh)max = maximum horizontal well oil flowrate
10 SPE 133436

(Qo) Max = Maximum oil flowrate, STB/day

re = reservoir outer radius, ft
rw = wellbore radius

References
1. Ahmed, T. 2001. Reservoir Engineering Handbook. 2nd ed, Gulf Professional Publishing, USA, P. 473- 515.
2. Bendakhlia, H. and Aziz, K. 1989. IPR for Solution-Gas Drive Horizontal Wells. Paper SPE 19823 presented at the 64th Annual Meeting in
San Antonio, Texas, October 8–11.
3. Borisov, J.P. 1964. Oil production using horizontal and multiple deviation wells, trans. J. Strauss. Bartlesville, Oklahoma: R&D Library,
Phillips Petroleum Co. (1984).
4. Cheng, A.M. 1990. IPR for Solution Gas-Drive Horizontal Wells. Paper SPE 20720 presented at the 65th Annual SPE meeting held in New
Orleans, September 23–26.
5. Fetkovich, M.J. 1973. The Isochronal Testing of Oil Wells. Paper SPE 4529 presented at the SPE 48th Fall Meeting Las Vegas, Nevada.
6. Giger, F.M., Reiss, L.H., and Jourdan, A.P. 1984. The Reservoir Engineering Aspect of Horizontal Drilling. Paper SPE 13024 presented at
the SPE 59th Annual Technical Conference and Exhibition, Houston, Texas, Sept. 16–19.
7. Gue, B.G., Lyones, W.C., Ghalambor, A. 2007. Petroleum Production Engineering. Elsevier Science & Technology Books, P. 30-36.
8. Joshi, S. 1991. Horizontal Well Technology. Tulsa, OK: Penn Well Publishing Company.
9. Renard, G.I. and Dupuy, J.M. 1990. Influence of Formation Damage on the Flow Efficiency of Horizontal Wells. Paper SPE 19414
presented at the Formation Damage Control Symposium, Lafayette, Louisiana, Feb. 22–23.
10. Retnanto, A. and Economides, M.J. 1998. Inflow Performance Relationships for Solution-Gas Drive Horizontal Wells. Paper SPE 50659.
11. Vogel, J.V. 1968. Inflow Performance Relationship for Solution Gas Drive Wells. JPT, (Jan. 1968): 86–92; Trans., AIME, 243.