Report Title:

Application of Artificial Neural Network (ANN) for modelling gas-condensate two-phase

compressibility factor
Table of Content

Contents
Abstract...................................................................................................................................3
1 Introduction......................................................................................................................3
2 Artificial Neural Networks...........................................................................................493
Figure 1: A Biological Neuron.....................................................................................493
Figure 2: A Processing Element or Neuron.................................................................494
Figure 3: The Neural Network Architecture used in this Study...................................494
3 Methodology................................................................................................................494
3.1 The Neural Network Architecture.........................................................................494
3.2 Training..................................................................................................................495
3.3 Results....................................................................................................................495
Table 1: Data Range for Input Parameters in this study..............................................495
Figure 5: Regression analysis of the neutral network performance for Validation.....496
Figure 7: Neural Network Performance in terms of Mean Squared Error...................496
References..........................................................................................................................497
Abstract
Gas compressibility factor is an essential requirement for the determination of several natural gas properties. Formation
volume factor, density, compressibility and viscosity all require accurate knowledge of the gas compressibility factor.
Limited availability of experimental data makes it necessary to employ correlations to calculate the gas compressibility
factor. This study evaluates the usefulness of Artificial Neural Networks (ANN) as an alternative to published correlations in
predicting gas compressibility factor. The ANN model correlates gas compressibility factor as a function of reservoir
temperature and dew point pressure. This method is of particular importance in generating gas compressibility data for
natural gas reservoirs where no samples have been taken.

ANN was applied to the 40 raw data sets in the range of 105-226 oF and 2445 - 4843 psia for temperature, and pressure,
respectively. The ANN network was constructed using MATLAB. To develop the ANN model, the samples were divided
into three groups. One set contained 24 samples which was used to train the network, one set of 8 samples was used for
validation while and the remaining 8 samples were used as the test sets. The performance analysis of ANN showed that the
mean square error (MSE) was 4.0769E-6 while the R 2 value for test data was equal to 0.99607. The model was tested in the
range of 107-117 oF and 2495 - 2725 psia for temperature, and pressure, respectively. Gas compressibility factors obtained
from the Standing and Katz chart gave a MSE of 4.14E-05 while the ANN gave a MSE of 3.02243E-05 as compared to the
experimental data. The results showed that ANN is an effective and powerful tool in estimating gas compressibility factor.

1 Introduction
Natural gas is a subcategory of petroleum that gas describes the behavior of most gases at pressure and temperature
occurs naturally, and it is composed of complex conditions close to atmospheric. At moderate pressures and low
mixtures of hydrocarbons and a minor amount of temperatures, natural gas tends to compress more than the ideal gas law
inorganic compound. Natural gasses physical predicts. Thus, the gas compressibility factor is defined as the ratio of the
properties, and in particular, their variations with volume actually occupied by a gas at a given pressure and temperature to
pressure, temperature, and molecular weight are of the volume it would occupy if it behaved ideally. It is denoted by the
great importance in petroleum and gas engineering symbol z. Thus, gas compressibility factor is expressed as:
calculations (Heidaryan et. al., 2010).
z = Va/Vi (1)
The natural gas compressibility factor is a
requirement in most petroleum and natural gas where, Va is the actual volume of gas and V i is the
engineering calculations. Some of these ideal volume of the gas both at the same conditions
calculations are gas metering, gas compression, of temperature and pressure.
design of processing units, and design of pipeline
The real gas equation is then written as:
and surface facilities. Compressibility factor of
natural gases is also important in the calculation of pV = nzRT (2)
gas flow rate through reservoir rock, material
balance calculations, evaluation of gas reserves,
and reservoir simulations (Azizi et. al., 2010).

The gas compressibility factor is a measure of the

deviation of the gas from ideal behavior. The ideal
1
where, P is the pressure, V is the volume, T is the logic are recognized as artificial intelligence tools
absolute temperature, n is the number of kilo-moles for data analysis, modeling and optimization.
of the gas, and R is the gas constant. Despite the fact that such tools are relatively new in
petroleum engineering, they are becoming more
At different phases of field development, the significant due to their ability to extract
natural gas compressibility factor needs to be information from data and transforming
predicted at different conditions of temperature and information into knowledge.
pressure. This can be determined through pVT
analysis of fluid sample tests or can be calculated As the term implies, Artificial Neural Networks are
by using an Equations of State (EoS) if the fluid networks of highly interconnected neurons found in
composition is known. This information is often the brain of many living organisms. The
unavailable particularly at the early stage of field simultaneous activity of such neurons had been
development or needs to be verified, supported and proven to impart intelligence. The human brain
supplemented during the course of field consists of a large number of interconnected
development. Empirical correlations are then used biological neurons. Each biological neuron has
to estimate the petroleum fluid properties as a three main parts (Figure 1): dendrites, a cell body
function of the reservoir's readily available (soma) and an axon.
characteristics.

The most commonly used method to estimate the

gas compressibility factor is the chart provided by
Standing and Katz which is based on the Theory of
Corresponding States. The chart covers the range
of reduced pressure from 0 to 15, and the range of
reduced temperature from 1.05 to 3 (Standing and
Katz, 1942). This provides reliable estimates for
natural gas and can be extended to gas condensate
and mixtures. However, empirical correlations have
been developed for determining the gas
compressibility factor instead using existing tables Figure 1: A Biological Neuron
and charts. The correlations are divided into two
groups: direct e.g. Beggs and Brill correlation and Connections between any two biological neurons
iterative relations e.g. Hall-Yaborough correlation (axon to dendrite) are called synapses. Naturally,
(Azizi et. al., 2010). electro-chemical signals enter via dendrites and are
transferred to the soma, where they are summed to
The use of artificial neural networks (ANNs) in form a combined signal. Provided that the
Petroleum Engineering has been increased recently. combines signal is strong enough, the neuron is
The wide applicability of ANNs stems from their activated and the signal is strong enough, the
flexibility and ability to model linear and nonlinear neuron is activated and the signal is transmitted via
systems without prior knowledge of an empirical the axon to the surrounding biological neurons.
model. This gives ANNs an advantage over This mechanism allows synaptic gap strength to
traditional fitting methods for some chemical increase and the human brain to communicate,
applications. ANN eliminates the limitations of the think and learn. The basic building component of a
classical approaches by extracting the desired neural network is called an artificial neuron (Figure
information using the input data. Applying ANN to 2), also known as a processing element (PE),
a system needs sufficient input and output data designed to perform in a similar manner to a
instead of a mathematical equation. ANN can be biological neuron.
trained using input and output data to adapt to the
system. Also, ANN can be used to deal with the
problems with incomplete and imprecise input data
(Heydari et. al., 2006). Hence this article is on the
use of ANN to predict compressibility factors of
some gas reservoirs from Niger Delta in Nigeria.

2 Artificial Neural Networks

Artificial neural networks, together with other
techniques such as genetic algorithms and fuzzy

5
Figure 2: A Processing Element or Neuron whereas the latter is utilized to check model
performance. However, designing a neural network
Signals (in mathematical form) enter a neuron in such a fashion may lead to a model that over-
through connections, which are assigned adjustable fitting problem, a modified approach, cross
values, known as weights. Each signal is scaled up validation, should be used in which case the data is
by its weight. The scaled signal is summed and a divided into three representatives data sets: a
bias value may be added. The final stage is to training set to calibrate the model, a testing set to
present the resulting signal to a non-linear check model performance at different stages during
activation function which, in turn, produces calibration and a validation set for the final
response to be supplied to other neighbouring assessment of the model. The calibration phase
neurons on the next layer. continues as long as both training and testing set
errors decrease. Calibration should determinate
For a typical multi-layer neural network, neurons
when the testing set error increases, see Figure 3.
are arranged into three layer types (Figure 3): a
This approach results in a model that has captured
single input layer (number of neurons is number of
the underlying relationship contained in the
input parameters), a single output layer (designed
calibration data, rather than a model that has
with neuron for each output) and one or more
“memorized” the specific data points in the training
hidden layers. It has been shown that, a network
set.
with a single hidden layer can model any
continuous function if an adequate number of An ANN is completely defined when the following
weights are provided. Such models have to be characteristics are specified: the architecture, i.e.
calibrated, where the overall aim is to obtain the number of units; how these are connected and
optimal values of weights and biases by how the information flows; the mapping, i.e. how
minimizing the error between predicted and to determine the outputs of the network as a
desirable output, by subjecting the model with function of its inputs; and the learning algorithm,
typical relationships for the problem under study. which is closely related to the architecture and
specifies how to adjust the weights of the network
during the learning process.

3 Methodology
3.1 The Neural Network Architecture
The Feedforward Multilayer Perceptron ANN
trained with the backpropagation algorithm has
been used in this study. This architecture has the
following characteristics:

Figure 3: The Neural Network Architecture 1. The architecture is multilayer. One hidden layer
used in this Study has been added between the input and output
layers. The addition of the hidden layer gives
As mentioned above, like regression models, the network the capability of approximating any
ANNs require a set of experimental (numerical) nonlinear function to arbitrary precision;
data that characterizes the relationship between 2. The information flows in a feedforward fashion,
inputs and outputs of the problem under i.e. from the input to the output going through
investigation. However, one additional requirement the hidden layers and, typically, the units in the
for regression models is to define a functional form previous layer only;
that, to a high degree, best represents the 3. The learning algorithm is supervised. The
inputs/output relation (i.e. they are formula-driven backpropagation algorithm is derived when the
models). ANNs, being data-driven, attempt to steepest-descent method is used to minimize the
extract such relationships from data provided sum of the squares of the error between the real
without a priori knowledge of the underlying output and the output predicted by the network.;
functional form.
The ANN Toolbox of MATLAB® R2009b was
Unlike regression models, the available data for used in developing the model and simulation in this
ANN development are, generally, divided into at paper work.
least two subsets: training and validation. The
former subset is used for model calibration,

5
 set of 20% of the total samples was used for
3.2 Training validation. The R-squared value and MSE were
0.99959 and 4.077E-06 respectively. The
There are generally four steps in the training remaining 20% of the total samples was used as the
process: test sets. The outputs were also given in figure 6
and 7 which revealed 0.99607 for the R-squared
1. Assemble the training data
value and 4.780E-06 for the MSE.
2. Create the network object
3. Train the network Table 1: Data Range for Input Parameters in
4. Simulate the network response to new this study
inputs
The data are orientated in such a way of presenting
Data Input Min. Max.
them in rows into the network (Figure 3). The
For Training, Temp. ( F) 105 226
o
network is developed through Graphical User
Validation Pressure (psia) 2445 4843
Interface with feed-forward backpropagation type.
and Testing SG 0.59 0.8
Standard backpropagation is a gradient descent For Temp. (oF) 107 117
algorithm, as is the Widrow-Hoff learning rule, in Simulation Pressure (psia) 2495 2695
which the network weights are moved along the SG 0.6 0.6
negative of the gradient of the performance
function. Input vectors and the corresponding target
vectors are used to train a network until it can
approximate a function, associate input vectors
with specific output vectors. Properly trained
backpropagation networks tend to give reasonable
answers when presented with inputs that they have
never seen. Training function employed was
TRAINLM which updates weight and bias values
according to Levenberg-Marquardt optimization.
Meanwhile LEARNGM was used as adaptation
learning function and for effective performance the
number of neurons for the first (hidden) layer was
ten while the second layer was one (usually
dictated by the number of output) (Howard and
Mark, 2002).

Networks with biases, a sigmoid layer, and a linear

output layer are capable of approximating any
function with a finite number of discontinuities.
Hence the usages of TANSIG as the first transfer
function of the hidden layer and PURELIN as the
transfer function for the output layer.

3.3 Results Figure 4: Regression analysis of the neutral

ANN was applied to the 40 raw data sets of network performance for training
temperature, pressure and specific gravity (S.G) in
It could be noticed from the results obtained that
the range of minimum and maximum values given
performance reduced in this order: training,
in Table 1. The samples were divided into three
validation and test. This could be because the data
groups by the network itself, selecting the number
used during training is more and knowing-fully
randomly. One set contained 60% of the total
well that ANN is data obsessed. Hence, the better
samples which was used to train the network.
result in training section while that of validation
According to the performance analysis of the
and test results could be because of the network
training section (Figure 4 and 7) where R-squared
seeing few new set of data. However, all these
value was recorded to be 1.0 and means square
excellent performances of ANN could be attributed
error was 3.400E-10 which showed that the
to only one major reason. The ability of neural
network was able to generate output that fit in
network to adapt and change its structure based on
(Figure 4) the target values very well. The second

5
external or internal information that flows through
the network during the learning phase.

Afterward, the model was simulated with new set

of data in the range of 107-117 oF, 2495 - 2725 psia
and 0.6 for temperature, pressure and specific
gravity respectively (Table 1). The simulated
results were plotted (Figure 8) alongside
experimental data and those read from Standing
and Katz chart with new set of data.

Figure 6: Regression analysis of the neutral

network performance for Test

Gas compressibility factors obtained from the

Standing and Katz chart gave a MSE of 4.14E-05
while the ANN gave a MSE of 3.02243E-05 as
compared to the experimental data. The results
showed that ANN is an effective and powerful tool
in estimating gas compressibility factor.

Figure 5: Regression analysis of the neutral

network performance for Validation

Figure 7: Neural Network Performance in terms

of Mean Squared Error

5
 Saputelli L., Malki H., Canelon J. and Nikolaou M.
(2002), “A Critical Overview of Artificial
Neural Network Applications in the Context of
Continuous Oil Field Optimization”, SPE
77703, 29 September – 2 October 2002.

Figure 8: Comparison of Experimental and

Predicted Compressibility Factor

4 Conclusions
The Artificial Neural Network developed in this
study has been shown to be a powerful instrument
for predicting compressibility factor most
especially when other data are not available to be
used in existing models. It correlates gas
compressibility factor as a function of reservoir
temperature and dew point pressure.

It also predicts gas compressibility factor

comparably to Standing and Katz chart, thus
providing accurate estimates without recourse to
the use of charts. It predicted compressibility factor
with Mean Square Error of 3.02243E-05.

References
Azizi, N., Behbahani, R. and Isazadeh, M. A.
(2010), “An Efficient correlation for calculating
compressibility factor of natural gases”,
Journal of Natural Gas Chemistry, 19 (2010)
642–645.
Bahadori, A., Mokhatab, S. and Towler, B. (2007),
“Rapidly Estimating Natural Gas
Compressibility Factor”, Journal of Natural
Gas Chemistry, 16 (2007) 349–353.
Heidaryan, E., Moghadasi, J. and Rahimi, M.
(2010), “New correlations to predict natural gas
viscosity and compressibility factor”, Journal
of Petroleum Science and Engineering.
Heydari, A., Shayesteh, S. and Kamalzadeh, L.
(2007), “Prediction of Hydrate Formation
Temperature for Natural Gas using Artificial
Neural Network”, Oil and Gas Business
Journal.
Howard D. and Mark B. (2002), “Neural Network
Toolbox: For Use with MATLAB”, The
MathWorks, Inc. 3 Apple Hill Drive, Natick,
MA 01760-2098.

View publication stats