Development of an Adaptive Neuro-Fuzzy based Model for Prediction

of Minimum Miscibility Pressure

A. Ameri a , Ameri@modares.ac.ir
M. Vafaie Seftie a, , vafaiesm@modares.ac.ir
S.A. Mousavi Dehghani b, mousavisa@ripi.ir
a

Department of Chemical Engineering, Tarbiat Modares University, Tehran, I.R. Iran, P.O. Box. 14115-4838
b

Research Institute of Petroleum Industry, NIOC, RIPI, Tehran, I.R. Iran, P.O. Box. 18745-4163

Abstract
In this paper, a neuro-fuzzy hybrid approach was used to construct a CO2 MMP
predicting system during design a gas injection project. In particular, we used an adaptive
network-based fuzzy inference system (ANFIS) to build a prediction model for reservoir
management. In neuro-fuzzy inference system, zero order Sugeno-type inference
technique was used to perform approximate reasoning of fuzzy input variables. In
addition, hybrid learning algorithm, combining back propagation learning and linear
least-squares estimator, was preferred for the adaptation of free parameters.
Consequently, neuro-fuzzy model was compared with results obtained using multiple
linear regression methodology in addition to other conventional models to make
comparison among different techniques. The results demonstrate that the ANFIS can be
applied successfully and provide high accuracy and reliability for MMP forecasting.
Keywords: Gas injection; Minimum miscibility pressure; Enhanced oil Recovery;
Neuro-fuzzy model
1. Introduction
Miscible gas injection is among the most widely used enhanced oil recovery
techniques, and its applications are increasingly visible in oil production worldwide. An
important concept associated with the description of miscible gas injection processes is
the minimum miscibility pressure (MMP). At this pressure, the injected gas and the initial
oil in place become multicontact miscible, and the displacement process becomes very
efficient.
The rationale behind the determination of MMP for a particular miscible gas injection
project is that there is a trade of between achieving high oil recovery and reducing
production costs. If the injection pressure is too low, the displacement would still be two

Corresponding author. Tel.: +98 21 88011001, Ext. 3979; fax: +98 21 88005040
E-mail address: vafaiesm@modares.ac.ir

1
www.petroman.ir

phase immiscible, and therefore the local displacement efficiency would be below the
desired level. If the pressure is too high, although the displacement would become
multicontact miscible, and the oil recovery would reach the desired level, the cost of
pressurizing the injected gas would be larger than necessary. Hence an optimal pressure
has to be found, and that pressure is MMP.
From an experimental point of view, the MMP is routinely determined by slim tube
displacements, Rising bubble approaches, and Vanishing Interfacial Tension technique.
Because such experiments are very expensive and time-consuming, searching or
developing a high precision mathematical determination of the gas-oil MMP is usually
requested.
In this paper, the advantage of a neuro-fuzzy computing technique in modeling the
prediction of minimum miscibility pressure in a gas injection process was examined. The
prediction begins by generating the fuzzy rules that define the relationship between the
input/output data using grid partitioning. Then, the optimization of the fuzzy rule set was
done with an adaptive neuro-fuzzy inference system (ANFIS). It has been well
recognized that the MMP for CO2 in a reservoir depends on reservoir temperature, oil
composition, and CO2 purity. For CO2 MMP modeling, the reservoir temperature,
molecular weight of oils plus fraction (here C5+), the ratio of volatile to intermediate oil
fraction, and pseudocritical temperature of injected gas were selected as input variables.
The data used for developing the neuro-fuzzy model are from differnt literature sources
[1-6]. The performance of the neuro fuzzy model was compared to the multiple linear
regression analysis and other conventional methods.
2. System modeling
The modeling process based on ANFIS can broadly be classified in three steps:
2.1. Step 1: system identification
The first step in system modeling is the identification of input and output variables
called the systems variables. Then fuzzy IFTHEN rules based on TakagiSugenoKang
(TSK) model [7,8] are formed, where antecedents are defined by a set of non-linear
parameters and consequents are either linear combination of input variables and constant
terms or may be constants, generally called, singletons.
2.2. Step 2: determining the network structure
Once the input and output variables are identified, the neuro-fuzzy system is realized
using a six-layered network as shown in Fig. 1. The input, output, and node functions of
each layer are explained in the subsequent paragraphs.
2.2.1. Layer 1 (input layer)

Each node in layer 1 represents the input variables of the model identified in step 1.
This layer simply transmits these input variables to the fuzzification layer.
2.2.2. Layer 2 (fuzzification layer)

The fuzzification layer describes the membership function of each input fuzzy set.
Membership functions are used to characterize fuzziness in the fuzzy sets. The output of
each node i in this layer is given by Ai ( xi ) where the symbol A (x) is the membership

2
www.petroman.ir

function. Its value on the unit interval [0,1] measures the degree to which element x
belongs to the fuzzy set A, xi is the input to node i and Ai is the linguistic label for each
input variable associated with this node.

Fig. 1. A typical ANFIS structure [1].

Each node in this layer is an adaptive node, that is, the output of each node depends
on the parameters pertaining to these nodes. Thus the membership function for A can be
any appropriate parameterized membership function. The most commonly used
membership functions are triangular, trapezoidal, Gaussian, and bell shaped. Any of these
choices may be used. The triangular and trapezoidal membership functions have been
used extensively, especially in real-time implementations, due to their simple formulas
and computational efficiency. However, since these membership functions are composed
of straight-line segments, they are not smooth at the corner points specified by the
parameters. Though the parameters of these membership functions can be optimized
using direct search methods but they are less efficient and more time consuming [9,10].
Also, the derivatives of these functions are not continuous so the powerful and more
efficient gradient methods cannot be used for optimizing their parameters. Gaussian and
bell shaped membership functions are becoming increasingly popular for specifying
fuzzy sets as they are nonlinear and smooth and their derivatives are continuous. Gradient
methods can be used easily for optimizing their design parameters. Thus in this model,
we used bell shaped fuzzy membership functions. The bell or generalized bell (or gbell)
shaped membership function is specified by a set of three fitting parameters {a, b, c} as
b
A ( x) = 1 / 1 + (( x c) / a) 2
(1)

The desired shape of gbell membership function can be obtained by proper selection of
the parameters. More specifically, we can adjust c and a to vary the center and width of
the membership function, and b to control the slope at the crossover points. The
parameter b gives gbell shaped membership function one more degree of freedom than
the Gaussian membership function and allows adjusting the steepness at crossover points.
The parameters in this layer are referred to as premise parameters.

3
www.petroman.ir

2.2.3. Layer 3 (inference layer)

The third layer is the inference layer. Each node in this layer is a fixed node and
represents the IF part of a fuzzy rule. This layer aggregates the membership grades using
any fuzzy intersection operator which can perform fuzzy AND operation [11]. The fuzzy
intersection operators are commonly referred to as T-norm (triangular norm) operators.
Most frequently used T-norm operators are min or product operators. For instance
IF x1 is A1 AND x2 is A2 AND x3 is A3 THEN y is f (x1, x2, x3)
where f (x1, x2, x3) is a linear function of input variables or may be a constant. The output
of ith node is given as
(2)
wi = A ( x1 ). A ( x2 ). A ( x3 )
1

2.2.4. Layer 4 (normalization layer)

The ith node of this layer is also a fixed node and calculates the ratio of the ith rules
firing strength in inference layer to the sum of all the rules firing strengths
wi =

wi

(3)

w1 + w2 + ... + wR

where i = 1, 2, . . ., R and R is total number of rules. The outputs of this layer are called
normalized firing strengths.
2.2.5. Layer 5 (output layer)

This layer represents the THEN part (i.e., the consequent) of the fuzzy rule. The
operation performed by the nodes in this layer is to generate the qualified consequent
(either fuzzy or crisp) of each rule depending on firing strength. Every node i in this layer
is an adaptive node. The output of the node is computed as
Oi = wi fi
(4)
where wi is a normalized firing strength from layer 3 and fi is a linear function of input
variables of the form ( pix1 + qix2 + ri), where { pi, qi, ri} is the parameter set of the node
i, referred to as consequent parameters or f may be a constant. If fi is linear function of
input variables then it is called first order Sugeno fuzzy model and if fi is a constant (as in
our present model) then it is called zero order Sugeno fuzzy model. The consequent can
be a linear function as long as it appropriately describes the output of the model within
the fuzzy region specified by the antecedent of the rule. But in the present case, the
relationship between the input variables (reservoir temperature, oils plus fraction
molecular weight, volatile oil fraction to intermediate oil fraction, and injected gas
pseudocritical temperature) and output (MMP) is highly non-linear. In Sugeno model,
consequents can be taken as singletons, i.e. real numbers without losing the performance
of the system.
2.2.6. Layer 6 (defuzzification layer)

This layer aggregates the qualified consequents to produce a crisp output. The single
node in this layer is a fixed node. It computes the weighted average of output signals of
the output layer as
O=

O = w f
i

wf
w
i

(5)

4
www.petroman.ir

2.3. Step 3: learning algorithm and parameter tuning

The ANFIS model fine-tunes the parameters of membership functions using either the
back propagation learning algorithm or hybrid learning rule [12]. Back propagation
algorithm is an error-based supervised learning algorithm. It employs an external
reference signal, which acts like a teacher and generates an error signal by comparing the
reference with the obtained response. Based on error signal, the network modifies the
design parameters to improve the system performance. It uses gradient descent method to
update the parameters. The input/output data pairs are often called as training data or
learning patterns. They are clamped onto the network and functions are propagated to the
output unit. The network output is compared with the desired output values. The error
measure EP, for pattern P at the output node in layer 6, may be given as
1

E =
P

(T O6 )
P

(6)

where TP is the target or desired output and O6P the single node output of defuzzification
layer in the network. Further, the sum of squared errors for the entire training data set is
E = E =
P

1
2

(T

O6 )
P

(7)

The error measure with respect to node output in layer 6 is given by delta ( )
=

= 2(T O6 )

O6

(8)

This delta value gives the rate at which the output must be changed in order to minimize
the error function. Since the output of adaptive nodes of the given adaptive network
depends on the design parameters so the design parameters must be updated accordingly.
Now this delta value of the output unit must be propagated backward to the inner layers
in order to distribute the error of output unit to all the layers connected to it and adjust the
corresponding parameters. The delta value for layer 5 is given as
E
O5

E O6

(9)

O6 O5

Similarly, for any Kth layer, the delta value may be calculated using the chain rule as
E
OK

E OK +1

(10)

OK +1 OK

Now, if is a set of design parameters of the given adaptive network, then

E

E O

(11)

O P

where P is the set of adaptive nodes whose output depends on . Thus, update for the
parameter is given by
=

(12)

where is the learning rate and may be calculated as

=

(E / )

(13)

5
www.petroman.ir

where k is the step size. The value of k must be properly chosen as the change in value of
k influences the rate of convergence.
3. Implementation
We have implemented our model using a MATLAB [13] code. The system is first
designed using Sugeno Fuzzy Inference System. It is a four inputone output system. The
input variables are reservoir temperature TR; the oils plus fraction molecular weight
(here MWC5+), ratio of volatile oil fraction (Xvol) to intermediate oil fraction (Xint), and
pseudocritical temperature of the injected gas. Molar average pseudocritical temperature
of injected gas is given by:
n
(14)
T =
wT
cm

ci

i =1

where wi is the molar fraction of component i; Tci is the critical temperature of component
i, and Tcm is the injected gas pseudocritical temperature.
The input parameters are represented by fuzzy sets (or linguistic variables). We have
chosen gbell shaped membership functions to characterize these fuzzy sets. Here, two
membership functions were assigned to each input. These membership functions map
each element of input space to a membership grade (or membership value) between 0 and
1. The membership functions are then aggregated using T-norm product operator to
construct fuzzy IFTHEN rules that have a fuzzy antecedent part and constant
consequent.
After construction of fuzzy inference system, the model parameters are optimized
using ANFIS. The network structure consists of 55 nodes. The total number of fitting
parameters is 40, of which 24 (four input with two membership functions and three
adjustable parameters, 4 2 3 = 24 ) are premise and 16 (four input with two
membership functions, 2 2 2 2 = 16 ) are consequent parameters. The total number of
rules is 16. Hybrid learning rule is used to train the model according to input/output data
pairs. Out of the total 55 input/output data sets, 44 data pairs were used for training the
model. The model was trained for 50 epochs with step size of 0.01 and error tolerance of
0%. To validate the model 11 data sets were used for testing purpose.
3.5. Multiple linear regression model development
Multiple linear regression analysis is a method used to model the linear relationship
between a dependent variable and one or more independent variables. The objective of
the multiple linear regression analysis is to determine the values of the parameters of the
regression equation and then to quantify the goodness of fit in respect of the dependent
variable. The details of the multiple linear regression analysis can be seen in Hocking
[14]. The general form of multiple linear regression is
Y = b0 + bi X i
(15)
where bo is the intercept, bi are the regression coefficients of the descriptor Xi and Y is the
predicted value. Multiple linear regression analysis was developed and tested with the
same data sets used for the neuro-fuzzy model, and the regression equations developed
were referred to as trained models. Then the predictive ability of the model was also
validated with the same data sets used to test the neuro-fuzzy model, thus making the

6
www.petroman.ir

result comparable. The validated model is indicative of the model capability to predict
MMP since the testing data are independent of the data used for model development.

4. Results and discussion

4.1. Performance of the developed neuro-fuzzy model

The MMP data were divided into two data sets, consisting of training and validation
test data. In the training phase, a larger part of the data was used to train the network and
the remaining data were used in the validation phase.
Fig. 2(a) present the error analysis from neuro-fuzzy model based on the training
data. Table 1 shows the outputs of statistical analysis for calibration results from the
neuro-fuzzy, statistical and GA-based models for MMP forecasting. It is indicated that
the developed neuro-fuzzy model has lower calibration errors than those developed by
Alston et al. [1], and Emera and Sarma [4]. In detail, in the training phase, the neurofuzzy model improved the statistical model forecast of about 69.9% and 56% reduction in
RMSE and MAPE values, respectively. In addition, improvements of the forecast results
regarding the correlation coefficient (R) and standard deviation (SD) values during the
training phase were approximately 7% and 56.86%, respectively. On the other hand, the
neuro-fuzzy model improved the GA-based model forecast of about 37% and 31.94%
reduction in RMSE and MAPE values, respectively. Also, for GA-based model,
improvements of the forecast results regarding the correlation coefficient (R) and
standard deviation (SD) values during the training phase were approximately 3.15% and
39.89%, respectively.
Table 1: Statistical analysis for calibration results from
the neuro-fuzzy, statistical and GA-based models.
Method
Neuro-fuzzy (This work)
Alston et al., [1] (statistical)
Emera and Sarma [4] (Genetic Algorithm)

MAPE

RMSE

SD(%)

3.58
8.14
5.26

0.56
1.86
0.89

4.49
10.41
7.47

0.98
0.91
0.95

In addition, in the validation phase as seen in Fig. 2(b), the values with the neurofuzzy model prediction were able to produce a good forecast. Table 2 shows the results of
error analyses for prediction outputs from developed neuro-fuzzy model and the
statistical and GA-based models. It is indicated that outputs from neuro-fuzzy model are
more accurate than those from the models of Alston et al. [1] and Emera and Sarma [4].

7
www.petroman.ir

Fig. 2: (a) The measured versus neuro-fuzzy simulated MMP values within 95% accuracy,
(b) The measured versus neuro-fuzzy predicted MMP values within 95% accuracy

Table 2: Error analysis for prediction outputs

Method
Neuro-fuzzy (This work)
Alston et al., [1] (statistical)
Emera and Sarma [4] (Genetic Algorithm)

MAPE

RMSE

SD(%)

4.00
12.57
5.86

0.61
2.47
1.07

5.23
12.68
8.38

0.98
0.81
0.92

In the validation phase, the neuro-fuzzy model improved the statistical model forecast
of about 75.3% and 68% reduction in RMSE and MAPE values, respectively. In addition,
improvements of the forecast results regarding the correlation coefficient (R) and
standard deviation (SD) values during the validation phase were approximately 21% and
58.75%, respectively. On the other hand, the neuro-fuzzy model improved the GA-based
model forecast of about 43% and 31.6% reduction in RMSE and MAPE values,
respectively. Also, for GA-based model, improvements of the prediction results regarding
the correlation coefficient (R) and standard deviation (SD) values during the training
phase were approximately 7.7% and 37.5%, respectively. The major advantage of the
neuro-fuzzy approach was the ability to capture well the major and minor trends in the
MMP series while the two other models failed in producing good results. Thus the results
indicate that the neuro-fuzzy model is able to identify the events for which it was
designed although the extent to which this model can generalize its ability to forecast
events was not included in the training process.
4.2. Comparison with multiple linear regression analysis
In order to show the potential of the neuro-fuzzy model, the forecast result of the
multiple linear regression model is also presented. The multiple linear regression model
was developed and tested with the same data sets used for the neuro-fuzzy model. The
regression equations that were developed are referred to as trained models and the

8
www.petroman.ir

predictive ability of the model was also validated with the same data sets used to test the
neuro-fuzzy model. The scatter plot in Fig. 3 provides comparisons of the measured CO2
MMP levels with provided by multiple linear regression model. The comparison of the
prediction accuracy using neuro-fuzzy and multiple linear regression analysis is
summarized in Table 3. It can be observed from Table 3 that the neuro-fuzzy model
performed better during both training and validation, and it outperforms multiple linear
regression in terms of all the standard statistical measures.
Table 3: Forecasting performance of neuro-fuzzy and
multiple linear models during the training and the validation stages
Performance index
R
SD
RMSE
MAPE

Neuro-fuzzy
Training Validation
0.96
4.49
0.56
3.58

0.97
5.23
0.61
4.00

Multiple linear regression

Training
Validation
0.97
9.16
1.00
7.06

0.92
9.34
1.37
7.94

Fig. 3: (a) The measured versus multiple linear regression simulated MMP values within 95% accuracy,
(b) The measured versus multiple linear regression predicted MMP values within 95% accuracy

5. Conclusions
Miscible gas flooding is widely employed for improving or enhancing oil recovery
for many oil reservoirs. A key parameter used for assessing the applicability of the
process for a reservoir is the minimum miscibility pressure. Therefore, accurate
prediction of minimum miscibility pressure is of the utmost importance.
An attempt was made in this study to investigate the application of a neuro-fuzzy
concept for prediction of MMP in a gas injection process. In this study, the neuro-fuzzy
approach is for the first time used to predict MMP. The MMP data derived from literature
were employed to train and test the models. The model was successfully applied to both

9
www.petroman.ir

pure and impure CO2 streams. In the case of impure CO2 stream, molar average
pseudocritical temperature was used as a statistic to show the level of contaminates that
exists in the CO2 stream. As a comparison, a multiple linear regression analysis was also
examined using various statistical indices in addition to the other statistical and genetic
algorithms approaches. The comparison of the prediction accuracies of the neuro-fuzzy
and other methods indicated that the neuro-fuzzy approach was more accurate in
predicting MMP. Thus, the results of this study suggest that the neuro-fuzzy model is
more reliable than other conventional methods for predicting MMP. Especially, under
conditions with limited field information, the neuro-fuzzy approach could produce a
higher accuracy than other forecasting methods.
References
[1] Alston, R.B., Kokolis, G.P., James, C.F., CO2 minimum miscibility pressures: a correlation
for impure CO2 streams and live oil systems, SPE J., 268274 (April, 1985).
[2] Emera, M.K., Sarma, H.K., Use of genetic algorithm to predict minimum miscibility pressure
between flue gases and oil in design of flue gas injection project, Paper SPE 93478 Presented
at the 14th SPE Middle East Oil & gas show and conference, Bahrain, 12-15March, (2005).
[3] Dong, M., Huang, S., Srivastava, R., Effect of solution gas in oil on CO2 minimum miscibility
pressure, J. Can. Pet. Technol. 39 (11), 53 61, (2000).
[4] Emera, M.K., Sarma, H.K., Use of genetic algorithm to estimate CO2oil minimum
miscibility pressurea key parameter in design of CO2 miscible flood, J. Pet. Sci. Eng. 46,
37 52, (2004).
[5] Rathmell, J.J., Stalkup, F.I., Hassinger, R.C., A laboratory investigation of miscible
displacement by CO2, Paper SPE 3483 Presented at the SPE 1971 annual fall meeting, New
Orleans, Oct. 11 12, (1971).
[6] Yellig, W.F., Metcalfe, R.S., Determination and prediction of CO2 minimum miscibility
pressures. J. Pet. Technol., 160 168 (Jan., 1980).
[7] Sugeno, M., Kang, G.T., Structure identification of fuzzy models, Fuzzy Sets Syst. 28, 1533,
(1988).
[8] Takagi, T., Sugeno, M., Fuzzy identification of systems and its applications to modelling and
control, IEEE Trans. Syst. Man Cyb., 15, 116132, (1985).
[9] Rao, S.S., Engineering Optimization-Theory and Practice, Wiley, New Delhi, (1998).
[10] Himavathi, S., Umamaheshwari, B., New membership functions for effective design and
implementation of fuzzy systems, IEEE Trans. Syst. Man Cyb., 31, 717723, (2001).
[11] Klir, G.J., Yuan, B., Fuzzy Sets and Fuzzy Logic: Theory and Applications, Prentice-Hall,
Englewood Cliffs, NJ, (1995).
[12] Jang, J.S.R., Sun, C.T., Mizutani, E., Neuro-fuzzy and soft computing: A computational
approach to learning and machine intelligence, Prentice-Hall, New Jersey, (1997).
[13] Fuzzy Logic Toolbox for use with MATLAB1, The Math-Works Inc. USA, (2000).
[14] Hocking, R.R., Methods and Applications of Linear Models-Regression and Analysis of
Variance, Wiley, New Jersey, (2003).

10
www.petroman.ir