Professional Documents
Culture Documents
1
Mohammed Alsharif Emhemmed (Member Of Petroleum Engineering Department, Sebha University,
2
Libya), Ir. ZUHER SYIHAB, PhD (Member of Petroleum Engineering stuff, Institut Teknologi Bandung,
Bandung 40132, Indonesia)
Abstract
EOR has a lot of methods and every method has validation to use it. One of
those methods is the CO2 injection and it is the first largest contributor to
global enhanced oil recovery. injection can be classified into two main
types: continues gas injection (CGI) and water alternating gas injection
(WAG). The objective of this research is to build a predictive model for WAG
injection by use the artificial neural network (ANN) and this model has ability
to give the same outputs that obtained from the WAG process simulator.
GEM-CMG simulator was used to valid the base case and check is work
probably, additional WINPROP-CMG was used to create variation PVT data
and CMOST-CMG used to do a sensitivity analysis for variation inputs and
provide outputs for data base of ANN. The numerical simulator uses 12 input
variables, including field properties and design parameters. Examples of
those inputs are porosity, permeability, gas rate,water rate, reservoir area,
reservoir depth, reservoir pressure, reservoir temperature, oil viscosity, oil API
and reservoir thickness. The same of those parameters will be the input data
for the input layer in the artificial neural network, also the output from the
simulator will be the output members of the ANN. ANN is not necessary to
use it during the petroleum process; ANN used just as a faster way for the
prediction. ANN is able to build just mathematical transform uses interconnect
sets of neurons that is able to adapt itself depending on the data fed to the
system. ANN should be learning to find the relationship between the input and
output by training, modify and testing the training parameters, this is called the
training process. If the minimum error has obtained that means the ANN
successful, then the ANN can predict the output just by giving the input
parameters. In this study, the minimum error in this paper is 7 %.
Corresponding Author:
Contact: mohammed_Samba@yahoo.com
INTRODUCTION
Reservoir simulation models are used by oil and gas companies in the
development of the fields. Where reliable model of a field is often time-
1
consuming and expensive; models are typically constructed only where large
investment decisions are at stake. Currently, a constructed predictive model
can give us an initial estimation, forecasting reservoir performance for a
variety of operating conditions, also the crude oil makes a major contribution
to the world economy today, where the attention is being paid to Enhanced Oil
Recovery (EOR) techniques for recovering more oil from the existing oil fields.
EOR has a lot of methods but this study included CO 2 WAG injection.
WAG injection has been widely applied since the late 1950s. Christensen, et
al 2001 have reviewed more than 60 fields and compared between those
fields, where WAG has been applied for different type of rock, different depth,
offshore and onshore successfully. Hence, this is our motivations to build a
predictive model for WAG by use Neural Network.
The concept of Neural Networks has been around since the early 1950s, The
neural network is a kind of artificial intelligence technology, it mimics
characteristics of biological neurons like the human processing system,
artificial neural network, acquire, store, and utilizing the knowledge by learning
the knowledge is embedded in the network that can be recalled in response
to the presented information (Zhong, et al 2001). The neural network has
been used in various petroleum engineering areas such as Petrophysics,
production, reservoir engineering areas etc. The neural network has been
successful approach to the time series prediction problems (Michael et al,
1990).
The main object of this paper is to predict model for WAG by using ANN
techniques and compare it with normal mathematical model, using reservoir
properties and the WAG design parameter as the input data for neural
networks by the variation of parameters and those parameters should be valid
to inputs, the second we have to generate data by using a numerical simulator
to provide the output layer for neural network and also compare the WAG
injection with conventional methods (water flood).
2
1. Training
These are presented to the network during the training, and the network is
adjusted according to its error. During the training they are another training
parameter play an important factor during developed neural network such as
learning rate and momentum. The learning rate is the training parameter that
control the size, while the momentum give a friction of the previous weight
update the current to the current weight .
2. Validation
These are used to measure network generalization, and halt training when
generalization stops improving.
3. Testing
These have no effect on training and so provide an independent measure of
network performance during and after training.
1. Input layer:
The input layer always 1 layer and the members of the input layer depend on
the number of input parameters.
2. Hidden layer:
The number of the hidden layer can be obtained by trial and error, where it
has been obtained a lot of instructions about how can be obtain the hidden
layer, but no one of those instructions can be followed, the number and
members of the hidden layer can obtain just by trial and error.
3. Output layer:
This represents the goals of neural network. It means the members of the
output layer depends how many goals that require to obtain.
They are neurons between those layers as shown in the figure 1. The main
purpose of those neurons (weights) to transport the information from the input
layer to the hidden layers and from the hidden layers to the output layer. The
complexity of real neurons is highly abstracted when modelling artificial
neurons. These basically consist of inputs (like synapses), which are
multiplied by weights (Strength of the respective signals), and then computed
by a mathematical function which determines the activation of the neuron.
Another function (which may be the identity) computes the output of the
3
artificial neuron (sometimes in dependence of a certain threshold). ANNs
combine artificial neurons in order to process information.
A (x, w) =
Where:
X; inputs parameter.
W; weights
The most common output function that used for ANN is sigmoid function also
known as a logistic function, is given by the following relationship:
Goal (x, y) =
Unswept area
Figure 2: The gravity effect for the gas injection Figure 3: The gravity effect of the water injection
WAG
Unswept area
Moreover, the WAG will improve the microscopic sweep efficiency through the
miscible CO2. Where the main purpose of the water injection to push the oil
out of the formation through the pressure by direct force, in the same time the
water injection will help to keep the reservoir pressure above the minimum
miscibility pressure, in the other hand, CO2 is able to mix with oil, that is going
to swell the oil. This mean the gravity force will be higher than the capillary
force, after some time the gas will vaporize the light component of the oil
where the oil viscosity will reduce and let the oil to move more easily through
the formation.
The main condition to get the miscibility is to keep the reservoir pressure
above the minimum miscibility pressure; the minimum miscibility pressure is
the lowest pressure at which the interfacial tension between a pair of fluids
vanishes and can be obtained by two ways:
2. Empirical correlations.
5
There are a lot of empirical correlations but in this study Yellings correlation
was used to calculate the minimum miscibility pressure, the MMP is function
of reservoir temperature, the proposed expression follows:
(Yelling,et al.,1980)
Methodology:
Start
Finish
Some of the parameters that depend on the Fifth SPE paper (The Fifth
Comparative Solution Project: Evaluation of Miscible Flood Simulators) such
7
as relative permeability and capillary pressure data. The another part of the
data depends on the design parameters such as gas injection rate and water
injection rate with the consideration the injection pressure less than fracture
reservoir pressure. The third part of data based on the field experiences,
where in this part should be making sure that the variations suitable for real
data. The field experience data, such as reservoir area. The fourth part of
data based on the screening criteria for Miscible WAG such porosity,
permeability, etc. The screening criteria data after SPE 130726 ( Recent
Developments And Updated Screening Criteria of Enhanced Oil Recovery
Techniques after A.Aladasani , et al., 2010 ). Table 1 shows the data from the
screening criteria:
PARAMETERS interval
Oil (API) 33-39
Viscosity (CP) 0.1-0.6
Porosity % 11-24 %
Permeability
130-1000
(MD)
Net thickness NC
Depth 7545-8887
Formation Sandston
temperature 194-253
Figure 5: Reservoir model for 1/8 five spot pattern. Figure 6: Reservoir model for 1/8 nine spot pattern.
8
Figure 7: Reservoir model for 1/8 inverse nine spot patterns
9
Table 3: The data included during creates PVT data for all the cases.
10
5. Water flood 1/8 nine spot patterns.
6. Water flood 1/8 inverse nine spot patterns.
Reservoir
porosity Permeability Water rate
area
=0.11,0.15,0.17 = 130,200,300,400,530, =239,192,1
=15,25,35
210 ,0.19,0.22,0.24 600,700,800, 1000 59
acre
Depth= 7545
120 , 7800, 8200
, 8887
30 ft Gas Water injection bbl
/d
injection(ft3/d)= =239. 8988359
114268. 991 , 191. 9190687
,1.37E+ 05 ,159. 9325573
,1.71E+05
11
that the curve of WAG injection will decline due to the pressure decline during
the gas injection, at sixth year noticed that the water scenario cannot displace
more oil, that is indicate the residual oil may be faced one of those
phenomena:
Table 4. Lists the input parameters that were used for base template:
12
Figure 10: comparison of recovery factor for miscible WAG and water flood for base case.
During the WAG scenario, when the gas bank contact with the oil bank, the
gas will condense inside the oil and oil will start to swell, the oil swelling during
the process of miscible CO2 flooding is main factor influencing the
effectiveness of this method to enhance oil recovery, since it will improve the
permeability of the rock when CO2 extracts the residual oil, swells it to let it
move. Thus, the gravity force will be greater than the capillary force, otherwise
the residual oil start to move but the oil droplets continues swelling until the
CO2 vaporize and rise with it the light components. During CO2 WAG
scenario, the main objectives to the water flood is to increase the reservoir
pressure to keep the reservoir pressure above the minimum miscibility
pressure, thus the injection rate will increase, and push the less viscous oil
forward to the production well, while during the gas injection the reservoir
pressure will decrease, thus injection rate will decrease, as shown in figure
13.
Figure 11: Comparison of cumulative Water Production for Water Injection versus CO 2 WAG Injection
for the base Case.
13
Figure 12: Comparison of cumulative injected Water injection for Water Injection Case versus CO 2
WAG Injection Case.
14
The Difference between the WAG And Water Flood
a) WAG reduced costs for gas lift, Injection of CO 2 gas, with the resulting
decrease water injection relative to the base forecast will cause reduced water
production over time. Figure 11 shows a comparison of water production with
and without CO2 WAG. A significant reduction in produced water with CO 2
WAG injection is observed.
b) Reduced future costs/penalties associated with handling of produced
water.
c) Increased capacity for oil, liquid capacity limited. If this capacity constraint
continues, we should see an increased capacity for oil with a reduction of
produced water. This will allow further acceleration of oil recovery as long as
the base forecast is total liquid constrained.
d) Acceleration of CO2 gas recovery, and Possible increased reserves due to
enhanced compaction drive.
e) The late breakthrough for WAG case as shown in the figure 11.
Figure 14 shows the sensitivity analysis for WAG ratio, where the WAG ratio
plays an important factor during the WAG injection. Where it could be
designed for two main reasons, the first reason based on the mobility control
and the second for economic purposes, in this figure has shown the increase
of the gas volume that means increase of oil recovery, the oil recovery
increase proportional with gas volume. During the WAG ratio 1:1.5 gave the
same oil recovery with WAG ratio 1:1.2, this is one of advantage of sensitivity
analysis for any fields. The WAG ratio sensitivity analysis will give us
economic prediction for the volume injection.
15
Figure 14: The sensitivity analysis for WAG ratio.
Example, if WAG has been applied for x field after some time found that gas
volume is not enough for next years, in this case we check how much the gas
volume that can be provide and fix it and it can be run sensitivity analysis for
water injection with variation water rate to obtain the best WAG ratio can be
applied for X field. Some fields have been run sensitivity analysis for WAG
ratio for economic purpose.
CMOST-CMG results
Some scenarios they have given gain by minus that is related to the gravity
segregation affect, this happened when the vertical reservoir permeability is
high and the high reservoir temperature support these phenomena. When the
gas vaporized the light component, this gas phase get high temperature,
where the increase of density inverse proportional with density where the high
temperature will help to rise the gas to the vertical permeability (upper
formation), thus decrease the horizontal flow velocity in the reservoir, the flow
path as shown in the figure 15. Also through the high permeability the
channelling of CO2 is reduced.
16
Figure 15. The gravity segregation affect
Figure 17 Shows how is the neural network fit the data. The whole X axis is
the prediction value for neural network and whole Y axis is the input value of
the neural network (the results from the simulator). Were R^2 for the water
flood is 0.95762 and R^2 for the oil recovery is 0.92554. Figure 18 Shows
input, output versus target where each point will focus on the direct line, any
point lies far from the main group indicate out layers, it may need to
investigate those as additional inside to the problem, the regression line
(colored lines) shows how the out puts centred around the target, also the
overall distance between regression line and direct line is summarize by R
value which should be 1.
17
Bias 1 Bias 2
WELL
RATIO I/P
Bias 3
area
(acre )
Bias 4
Depth (ft)
Porosity
Permeability WAG RF
(md)
Gas rater
(ft/d)
Water RF
Water rate
(bbl/d)
Reservoir
pressure
(psi)
Thickness
(ft)
Temperature
(f)
Oil viscosity
(cp)
Input layer
Oil API
Hidden layer
Out put
layer
Figure 16: The final architecture model of the artificial neural network.
18
Table 5: Summary of ANN results and mathematical model results.
19
Table 10: Represent the results of back propagation neural network for WAG
recovery results compared the results of simulator with ANN results.
20
The figure 19,20 represented some results of the validation data for the
forward ANN developed for all the job patterns it couldnt pick up all the
validation data in this figure because it will not be clear enough just has
drawn the first 50 sample. They all use three hidden layers in the architecture
with 12, 9 and 6 neurons respectively and a learning rate and momentum of
0.8.
Figure 19: Represents match data for recovery factor during WAG vs. Some number of validation data
Figure 20: Represents match data for recovery factor during water flood vs. Some number of validation
data
The flowing figures shows error % vs. the sample numbers for all the
validation data where the error % for WAG oil recovery was 1.84% while for
water oil recovery the error 0.84%.
Figure 21: Represents the oil recovery factor error % vs. Number of validation data during WAG.
21
Figure 22: Represents the oil recovery error % vs. Number of validation data during the water flood.
Where:
Well pattern: the ratio of the injection wells to the producing wells based on
the spot pattern unit less.
Area: reservoir area (acre).
Depth: reservoir depth (feet).
Permeability: (md).
Gas rate :( FT3/day).
Water rate: BBL/day
Oil viscosity: (centipoises)
API: oil density API
A= -3.60018, a=0.03066, b=0.03629, c=0.000334, d=0.0201, e=0.01101,
f=0.078559, g=0.02303, h=0.335316, j=3.472679, m=0.02043
22
Figure 23: Represents the WAG recovery factor error % vs. Number of validation data (mathematical
model)
The mathematical model that obtained during the water flood as following:
Where:
Figure 24: Represents the oil recovery factor error during water flood % vs. Number of validation data
(mathematical model)
23
Conclusion:
A predictive model has developed in compositional WAG floods. The
model with adjustable parameters in the homogenous reservoir, by use
two tools Artificial Neural Network and mathematical model.
The model that obtained by use artificial neural network has proven that it
is more accurate and accept more parameters than mathematical model
that obtained by regression.
WAG provides high recovery for the x filed when compared with
conventional methods due to high sweep efficiency.
Oil swelling plays important phenomena to delay the breakthrough due to
reduce the water relative permeability.
Increase water injection rate gave low RF due to the Flow Dispersion
Effects.
The most favourable conditions to happen gravity segregation are high
temperature and high vertical permeability.
References
24
25