Journal of Petroleum Science and Engineering 122 (2014) 69–82

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

journal homepage: www.elsevier.com/locate/petrol

A novel tool for designing well placements by combination of modified

genetic algorithm and artificial neural network
Tutuka Ariadji, Febi Haryadi, Irfan Taufik Rau, Prasandi Abdul Aziz, Rinaldy Dasilfa
Department of Petroleum Engineering, Institut Teknologi Bandung, Bandung 40132, Indonesia

art ic l e i nf o a b s t r a c t

Article history: Well placement optimization techniques that use reservoir simulations are currently taking advantage of
Received 5 December 2013 using the Genetic Algorithm method and involve the output of a reservoir simulation, which is the
Accepted 23 May 2014 hydrocarbon recovery, and thus, the technique requires running a reservoir simulation when finding a
Available online 24 June 2014
maximum value for the recovery. For a very large field of gas, a condensate reservoir would be very time
Keywords: consuming, and when there is only a limited amount of time for decision making, this approach would
well placement not be a sufficient technique. Of course, the conventional, traditional trial-and-error technique requires
genetic algorithm more effort. To address this very common challenge in field development planning, we propose the
artificial neural network concept of transferring the manual traditional technique into a novel tool technique that employs
Genetic Algorithm (GA), which can be used as a plug-in software application.
This paper employs a specifically formulated Genetic Algorithm method for applications in well
location optimization by introducing a newly proposed fitness function (objective function) that
was constructed from basic reservoir engineering properties, i.e., permeability, porosity, oil saturation,
pressure of reservoir, and thickness. Furthermore, this Genetic Algorithm method was then further
extended to consider the drainage radius, existing wells, existence of faults and multiple layers,
simultaneously. Hence, a software application has been developed that incorporates all of these concerns
into a rapid tool.
Reservoir modeling cases of oil and gas fields were used to test the proposed method, with the intention
of showing the rapidness of finding the well locations, and as an additional output, this approach could
yield a higher recovery than the previous technique, overall. The oil field is for cases in which multiple wells
penetrate multi-reservoirs typically and penetrate selected reservoirs in test cases. However, the gas field
application is for the case of horizontal well placements in which the direction and length are the optimized
parameters, which are optimized by employing an Artificial Neural Network (ANN) method for the length
optimization after having the best direction obtained from the Genetic Algorithm (GA) method. The
proposed method can give hydrocarbon recovery results in a much faster way and even better values than
the conventional method.
& 2014 Elsevier B.V. All rights reserved.

1. Introduction
Determining the locations of wells is one of the critical
problems in the exploitation of oil and gas fields, both in the
development of new fields and in the management of mature
fields. The process of determining the well locations is an optimi-
zation problem that involves gaining oil or gas as much as possible
by drilling a number of wells, with the total being as small as
possible. Until now, the determination of the locations of new
wells has usually been conducted by using a conventional trial-
and-error method, when placing coordinates at a geological or
E-mail address: tutukaariadji@tm.itb.ac.id (T. Ariadji).
reservoir model and then running with a simulator to obtain the
oil or gas recovery; this process requires experience, considerable
time, and a relatively high cost. Especially when addressing oil
fields that have a relatively large size, we are faced with the
possibility of a remarkably large number of solutions (locations of
wells), which causes the conventional methods to become inef-
fective and inefficient.
Furthermore, genetic algorithms in the oil industry world have
been applied in various areas, starting from the optimization
of pipe diameters, determining the distribution of the pressure,
the detection of corrosion, and so forth, including the issue of
determining the locations of the well. However, in determining
the locations of the well, in general, the Genetic Algorithm (GA)
has always been applied as an integrated unit with a reservoir

http://dx.doi.org/10.1016/j.petrol.2014.05.018
0920-4105/& 2014 Elsevier B.V. All rights reserved.
70 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82
simulator (Bittencourt and Horne, 1997; Dianati et al., 2002;
Montes and Bartolome, 2001). In such cases, the locations of the
wells proposed by the GA in each iteration will always be
evaluated by using a reservoir simulator. Of course, for large size
reservoirs or complex reservoirs, blind properties require consid-
erable time to run the simulation in a reservoir simulator. There-
fore, a different approach that can serve as an alternative solution
to the problem is needed, and a new method based on the
development of the Genetic Algorithm method is performed to
provide a much more effective and efficient method of solving the
problem.
This approach intention is to avoid running a reservoir simu-
lator as the current well placement method by employing basic
reservoir engineering properties and applying the genetic algo-
rithm as an evaluation judgment. This new proposed method is
basically transferring a manually placing best well coordinates at a
grid system of a reservoir model during a reservoir simulation
process into a computerized calculation of well placement based
on reservoir engineering data.
Thus, this study uses basic reservoir engineering data to evaluate
the well placements; and the GA is applied in a way that it is not an
integral part of the reservoir simulator. The reservoir simulator is
used at the end of the process, to validate the well locations that
were proposed by the GA. The results of the reservoir simulator are
then compared to those of the conventional methods.
A previous study (Ariadji et al., 2012) was still limited to the
determination of the locations of wells in a one-layer reservoir, but
in this study, this approach has been extended to the determina-
tion of the locations of multiple wells in a multi-layer reservoir.
Moreover, this approach also has been expanded to a horizontal
well direction and length optimization by combining the GA and
the Artificial Neural Network (ANN) methods.
After we obtain the results in this study, then we need to
validate it. All of the results in this study then will be validated by
using commercial reservoir simulator, i.e. ECLIPSE which is used
most commonly by reservoir engineers.
1.1. Objective of this study
The objective of this study is to develop a rapid tool that
employs the Genetic Algorithms (GA) and the Artificial Neural
Network (ANN) method to optimize multiple vertical wells and
the horizontal well placements. The proposed method requires the
same data as the manual conventional method for reservoir data,
i.e., porosity, permeability, thickness, saturation, and pressure in
the format of a reservoir model grid map, for developing the GA
objective functions. At this moment, this study does not involve
the presence of fault and stress distributions.
1.2. Methodology approach
The basic reservoir engineering properties of a reservoir model
in grid distributions have the value of permeability (k), porosity
(ϕ), current oil saturation (So), thickness (h), and current pressure
(P) as required input data, and these variables would be the
criteria of the optimization in finding the best well location. Based
on the reservoir engineering fundamental principles, as a logical
consecuence, for a vertical well placement case, the best location is
where all of the above parameters are at their highest values at a
certain coordinate, but typically, this location is almost impossible
to find. Thus, a logical consequence is that the multiplication of all
of the parameters would be the standard approach to comparing
each property parameter in the best way. In addition, in a multi-
reservoir field, there are criteria for the selection of reservoirs that
have potential to be perforated for production. Moreover, we
consider that the well drainage radius and the fault boundary would
be needed. On the other hand, for a horizontal well placement case,
two variables must be optimized, i.e., the direction and length of the
horizontal well segment. Consequently, such limited multiplication
of the reservoir engineering property approach is not sufficient,
and thus, we need an additional method to involve the second
variable, i.e., the horizontal well length, which can be accomplished
by applying the neural network approach to relate the input (the
length of the horizontal segment) and the output (the hydrocarbon
recovery) while using as limited an amount of data as possible. Thus,
we propose a combination of genetic algorithms and artificial neural
networks to automatically process all of the above considerations
and to optimize for the best well locations.
The proposed method employs the multiplication of the basic
reservoir engineering properties as a new objective function
for the genetic algorithm method. To validate the results of the
proposed method0 s best locations, a reservoir simulation is then
conducted to compare the results with those of the conventional
methods in terms of hydrocarbon recovery. The results of the
conventional methods are obtained directly from the previous
reservoir studies on the respective fields. The overall methodology
is performed as depicted in Fig. 1.
1.3. Description of the reservoir model for testing the method
This reservoir model is generated from a real oil field, namely,
X field, that has a formation of generally dominated by sandstones,
and the depositional environment of fluvio deltaic shallow marine.
This sandstone is very fine to very coarse grained, angular to
rounded grained, with poor to well sorting, hard but some parts
are not consolidated properly.
The X field has an average value of good permeability of 147 mD,
a medium porosity of 12.4%, a sufficient oil saturation of 35.5%, a thin
layer thickness of 1.86 m, and a reservoir pressure of 171 bar. The
reservoir characterization indicates that the permeability and oil
saturation are quite heterogeneous. The reservoir model has a total
number of layers of up to 12, in which the optimum locations of wells
for 2 conditions will be sought: the first condition is with all layers
perforated, and the second condition is with selective perforated
layers based on specified criteria. The model of the reservoir itself
has been scaled up into 283,912 grids. Fig. 2(a)–(c) shows the
distribution of the permeability, pressure, and oil saturation; grids
with dark colors have lower values, while grids with bright colors
have higher values.
In X Field, the depletion mechanism is dominated by the water
drive. In addition to the water drive, this field also has a gas cap,
which serves as its additional depletion mechanism.
1.4. Application of the genetic algorithm for simultaneous vertical
well placement
The Genetic Algorithm (GA) is an optimization method for obtain-
ing the optimum solution to a problem that is formulated in
terms of mathematical functions. Compared with other optimization
Data: Collecting data input of the
Reservoir Model reservoir for the GA method:
k, φ, So, h, & pressures
Developing
GA Method: Objective
Validating with the
Software use of conventional
Function, Algorithm,
Computer Coding and
Development reservoir simulation Combining with the ANN
(for the horizontal well)
Fig. 1. Research methodology.
 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82 71

Fig. 2. (a) Permeability distribution of X Field (Ariadji et al., 2012). (b) Pressure distribution of X Field (Ariadji et al., 2012). (c) Oil saturation distribution of X Field (Ariadji
et al., 2012).

algorithms, the GA is a robust algorithm, which means that these Generate

Fitness YES Best
algorithms are not easily trapped into a local optimum like some random initial Representation Max
population evaluation individuals
other optimization algorithms (Ozdogan and Horne, 2004). The GA
was first introduced by Holland in 1968 (Pham and Karaboga, 2000). NO

In general, the procedure in the GA was inspired by Darwin0 s theory

of evolution (Pham and Karaboga, 2000; Romero et al., 2000).
In Darwin0 s theory, natural law prevails, which means that
excellent individuals in a population will persist, whereas weak
individuals will become extinct and be replaced by better indivi-
duals. Thus, future generations are usually better than previous
generations. In between the generations, marriages between indivi-
duals occur, and out of these marriages come offspring (new
individuals) that are better and will live into the next generation.
To produce a better offspring, there is a selection process in which
individuals that excel tend to marry individuals that also excel.
Following the existence of new individuals, the weak individuals will
be replaced in the next generation. Among the generation alteration,
mutations can probably occur, i.e., changes in the structure of the
genes due to specific reasons. These mutations make the individuals0
characteristics change from their previous states. In general, these
mutations are not expected to occur. That scenario constitutes the
cycle of natural selection that occurs in general in the GA.
The overall procedure of the GA method is described in Fig. 3,
as follows.
The following are the step-by-step specifics of the procedure in
Fig. 3 (Pham and Karaboga, 2000).
1.5. Generating a random initial population
Determination of the initial population was performed ran-
domly. The individuals in the population are represented in binary
Generate Selection
new END
START
population
Mutation Crossover
Fig. 3. The general GA method procedure.
as chromosomes of those individuals (Mitchel, 1996). These
chromosomes are composed of genes that have a specific length.
In the GA, the chromosome length is called the bit length. The larger
the bit length of the individuals generated is, the denser (more) the
possible solutions. For continuous solution areas, the greater the bit
length, the more accurate the obtained solutions are.
The following (Fig. 4) is an example of a randomly generated
population, where the population is made up of 10 individuals that
have a bit length of 20.
In determining the locations of a well, the composition of the
genes of each individual represents the locations of the wells in
grid (x, y). The first 10 bits represent the genes0 abscissa x, whereas
the last 10 bits of the genes represent the ordinate y. For example,
for Individual 1 in Fig. 10 above, the first 10 bits in the gene, i.e.,
1011001100, will represent the abscissa x, and the last 10 bits of
the gene, i.e., 0000101011, will represent the ordinate y, and so
forth for the next individuals.
72 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82
Individual 1: 10110011000000101011
Individual 2: 11000100101011100001
Individual 3: 01101111110000010110
Individual 4: 10101010111110110000
Individual 5: 11100011110010000010
Individual 6: 00110010001110000000
Individual 7: 00110110110101001101
Individual 8: 11001000101101011111
Individual 9: 11111101001001001111
Individual 10: 11001000101010000001
Fig. 4. Examples of a randomly generated population.
Determination of the number of individuals within the popula-
tion is crucial. The larger the number of individuals (members of
the population) is, the larger are the estimates of the locations of
the wells, which means the greater is the probability of getting the
best individuals (the optimum solution). More individuals (a larger
population) means that there are more comparative opponents of
the best individuals in the previous generation, which means the
easier it is to obtain the best individuals. In other words, the larger
the population is, the larger is the opportunity to obtain the
optimum solution and the smaller the number of iterations that is
required. However, enlarging the population is not preferable in
terms of the computation. The larger the population is, the longer
the time that is needed at each iteration of the computation. Thus,
for a maximum (value) of the same iterations, the time that is
required by the GA method for a population with a larger number
of members is longer.
2. Representation
After a population is obtained, each individual is subsequently
transformed from a binary string into a decimal number. If n is the
bit length of an individual, then the values of x and y that are
previously a binary string will be transformed into the decimal
numbers x0 and y0 within the interval of [0, 2(n/2)  1] Pham and
Karaboga (2000).
Next, x0 and y0 , which have values that are in the interval of
[0, 2(n/2)–1], will be mapped into a grid (x, y) of the reservoir model
size (a,b). The mapping is performed by using the following
functions, given in Eqs. (1) and (2) (Pham and Karaboga, 2000):
ða  1Þ x0
x¼ ðn=2Þ
þ1
ð2 –1Þ
ðb  1Þ y0
y¼ þ1
ð2ðn=2Þ –1Þ
With these mappings, eventually the population that was pre-
viously represented with a binary string can be transformed into
the position (x, y) in the reservoir section.
3. Objective function evaluation (fitness evaluation)
Next, each individual in the population has a value with grid
coordinates (x, y). These coordinate values are candidates for the
locations of the well. Each candidate/individual will then be
evaluated by using an objective function. In this study, there are
three objective functions that are to be compared with one other,
as follows:
k i¼1
F 1 ¼ ∑ permeabilityðx; yÞ  oil saturationðx; yÞ  porosityðx; yÞ
i¼1
k
F 2 ¼ ∑ permeabilityðx; yÞ  oil saturationðx; yÞ
i¼1
 porosityðx; yÞ  pressureðx; yÞ ð4Þ
k
F 3 ¼ ∑ permeabilityðx; yÞ  oil saturationðx; yÞ
i¼1
 porosityðx; yÞ  pressureðx; yÞ  thicknessðx; yÞ ð5Þ
The objective functions will evaluate the desired layers; from first
layer to the k-th layers, the layers to be perforated are divided
into two scenarios. The first scenario is with all layers perforated,
and the second scenario is when the layers to be perforated are
required to meet the following reservoir engineering principles:
Principle 1: multiplication of the permeability and reservoir
thickness, ZC1, and
Principle 2: oil saturationZC2,
where C1 and C2 are constants that are determined according to
the best practices in the reservoir engineering.
Overall, there are 4 scenarios to be compared with the results
from the previous reservoir study. The four scenarios are as
follows: (1) Objective function F1 with all layers perforated;
(2) Objective function F2 with all layers perforated; (3) Objective
function F3 with all layers perforated; (4) One of the three
objective functions above but with a rule that only layers that
follow Principles 1 and 2 will be perforated.
The GA will maximize the objective function to obtain the
individuals that have the largest fitness value. The fitness value
itself is a result of the calculation of the value of the objective
function value.
4. Generating a new population
Following the evaluation of each member of the population, the
next step is to establish a new population. In forming this new
population (generation), there are 3 (three) steps to be taken,
namely, selection, crossover, and mutation (Pham and Karaboga,
2000).
5. Selection
The initial phase prior to crossover constitutes conducting a
ð1Þ selection. Selection in this sense can have two meanings. The first
meaning is natural selection, in which some of the best individuals
are selected to exist/live in the next generation (become a member
ð2Þ of the new population). In general, usually only one or two
individuals are selected. The second meaning is a matching
selection among individuals to perform the process of crossover.
In this second meaning, an individual that excels tends to be
matched to an individual that also excels, with the expectation
that they could obtain a superior offspring.
In the process of the matching selection among the individuals,
there is a process of random probability, specifically, the process
of selecting parent 1 and parent 2 to subsequently conduct the
process of crossover. Individuals that excel have a better chance to
be selected to be parents. To select the parents, the steps are as
follows (Pham and Karaboga, 2000).
Calculating the total fitness of the population:
N_pop
F total ¼ ∑ F i
ð3Þ
where N_pop is the size of the population.
 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82 73

Calculating the probability of the selection of each individual:

Fi
Pi ¼
F total
Calculating the cumulative probability for each individual:

i
Qi ¼ ∑ Pk
k¼0

4. Selecting parent 1 and parent 2

a. Generate random numbers in the range [0,1].
b. If Q i  1 o r r Q i , then select the i-th individual as parent 1.
c. Perform steps 4a. and 4b. again to obtain parent 2.
Fig. 6. Flow chart of the crossover process.
The physical consequent of this selecting criterion is that the
higher ΔQ (¼Qi–Qi  1) will be selected as parent. In other
words, a higher value of total fitness will have a higher
probability to be selected.
Table 1
Parameter used in this genetic algorithm applica-
tion study.
6. Crossover
Parameter Values
After parents 1 and 2 are obtained, the next step is as follows.
Maximum iteration 100
First, decide whether there is a crossover or not between
Number of individuals 500
parent 1 and parent 2. This decision can be made by generating Length of bit/chromosomes 20
a random number, r, in the range [0,1]. If the random number r is Crossover probability 0.97
lower than that of the crossover probability, then perform the Mutation probability 0.01
matching (do the next step). However, if r is greater than the
probability of the crossover, then no crossover will occur and there
will be no birth of new individuals (if r 4 P c , then stop until this
individuals, and the two best individuals are retained to exist in the
step only) (Pham and Karaboga, 2000).
next generation (the new population). The result is that eight new
The value of the probability of a crossover is within the range
individuals will be needed. Therefore, perform steps 1–3 four times.
[0,1]; the closer the value is to one, the larger is the probability for
For more detail, a crossover flow chart of the process can be
a crossover to occur; in contrast, the closer the value is to zero, the
seen in Fig. 6.
more unlikely the crossover is to occur. Usually, the probability
of the crossover is closer to one for a crossover to occur and to
produce a new individual. In this study, the probability of a
7. Mutation
crossover is 0.97.
In case crossover occurs, the next step is to determine the cut
Once the crossover is completed, the next process is mutation.
point (cp). A cut point cp is selected at the back of a gene that is
Mutations are changes in the genes that alter an individual0 s genes in
located closest to the integer that is greater than or equal to cp,
such a way that they are different from the genes of their parents. In
where cp ¼ rðL_bit  1Þ and L_bit is the length of the bit. Thus, the
GAs, mutation is performed by replacing one or more genes with the
cut point depends on components of random number r and the
opposite genes. In general, the probability of a mutation occurring is
length of the bit L_bit.
very low (in this study, the probability of the mutation is 0.01). Thus, in
After a cut point is selected, interbreeding is performed. In this
a population, genes that change are usually very limited in number; it
study, the first individual chromosome is a combination of genes
is very likely that mutations would not occur at all.
before the cut point of the first parents and those after the cut
The mutation procedure is performed through the following
point of the second parents. At the same time, the second new
steps (Pham and Karaboga, 2000): (1) Generate a matrix of
individual chromosome is a combination of genes before the cut
random numbers that has the size of the number of individual
point of the second parents and the genes after the cut point of the
times the bit length of each. (2) If the matrix consists of numbers
first parents.
whose values are below the probability that is required for the
For example, the sample chosen as the first and the second parents
mutations, then at the position of the member, mutations should
in a sequence are individuals 4–7 in Fig. 5, and 6 cp ¼0.7  18; as a
be conducted for the new population. If the position of the
result, the crossover process is described as follows:
mutation in the new population has the value of one, then replace
Repeat steps 1–3 several times until the required number of
it with a zero, and vice versa, i.e., if the position of the mutation in
new individuals is reached. Suppose that a population consists of 10
the new population is zero, then replace it with one. (3) Excellent
individuals that result from the previous generation and that exist
in the new population are not the object of the mutation. There-
Parent 1: 10101010111110110000 fore, if such individuals are mutated, then restore their excellent
chromosomes to their previous state.
Parent 2: 00110110110101001101
Once a mutation is complete, a new generation (population) is
New individual 1: 10101010110101001101 formed. Then, repeat the process of representation, objective func-
New individual 2: 00110110111110110000 tion, evaluation, and so on. This GA procedure is to be repeated until
the maximum number of iterations is met. The GA parameters that
Fig. 5. Example of the interbreeding of chromosomes in the crossover process. are used in this study are listed in Table 1.
74 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82
8. A new modified genetic algorithm procedure for
simultaneous vertical multi-well placement
Basically, the GA method is used to find a single optimum
solution (Ariadji et al., 2012). However, in the topic of this study,
we are not only looking for a single optimum solution (a first
optimum solution) but a second, third, fourth, and other optimum
solutions. When we are looking for only one location of a well, we
need to find only a single optimum solution. However, when we
look for a combination of locations of wells (more than one), we
are required to locate the best wells of the second, third, fourth
ranking, and so on. Therefore, it is necessary to modify the GA
to obtain not only a single optimum solution but many optimum
solutions.
The most major modification is made in the selection of
individuals who persist into the next generation (individual
selection). In the conventional GA as described by reference,
generally, only a few superior individuals are allowed to exist in
the next generation (usually one or two individuals only). In this
study, the GA is modified to enable new individuals (results of
crossover) to replace all of the older individuals that are of lower
quality. Obviously, the older individuals that are superior to new
individuals will not be replaced. In other words, the modified GA
makes the process of natural selection more objective than the
conventional GA, and, this elitism process makes individuals of the
new next generation equal to the n best individual number, where
n is the number of individual in a population. The modifications
made can be seen in the following flow chart. Fig. 7 described this
new GA procedure.
A modification that is equally important is the modification in
terms of computer programming. The above mentioned modifica-
tion takes a longer time at performing the iteration than the
conventional GA, but in fact, with some technical modifications, it
can be avoided; this modification can even make the GA converge
faster than the conventional GA.
There are 4 cases with respect to the objective function for the
newly developed GA, which are the following: (1) F1 with a
scenario of all layers perforated; (2) F2 with a scenario of all layers
perforated; (3) F3 with a scenario of all layers perforated; (4) A
function with a scenario of layers to be perforated in such a way as
to satisfy certain constraints.
The basic strategic concept of developing the objective functions
is to utilize the fundamental reservoir engineering parameters that
are used for reservoir simulation, i.e., data input in the grid system
for a reservoir simulator. The logical reason of the F1 objective
function is a base formulation that is constructed from the principles
of fundamental reservoir engineering as explained in the following.
Multiplication of porosity and saturation is more or less an oil
volume, and thus, the higher value of this multiplication results in
Fig. 7. Modified GA procedure.
higher volume of oil for the grids. Moreover, continuing operation by
multiplying with permeability to the previous multiplication term
would be involving the dynamic property concept that represent
how fast that oil volume could move. Thus, this is actually oil
volume movement capacity of each grid.
Continuing concern about extending the base objective func-
tion (F1) is to add a geological parameter that is significant and
distinguish from the previously involved parameters. Reservoir
thickness is a very important parameter representing volume of
reservoir and is independent to porosity, saturation and perme-
ability. Thus, inclusion of this parameter makes an aggregate
parameter of oil volume more complete than the base formulation
as well as the oil volume movement capacity, i.e., multiplication of
porosity  thickness  saturation  permeability (F2). Final attempt
is to comprehensively involve a dynamic parameter in addition to
permeability that is pressure of each grid to complete the flow
potential concept of objective function (F3).
The performance of all of the well locations that result from
this modified GA will be evaluated by using the reservoir simu-
lator, and then, they will be compared with the performance of the
well locations that are proposed by a previous reservoir simulation
study, using the conventional method of trial-and-error. For cases
1–3 above, the results would be compared with those of conven-
tional methods that used the scenario that has all of the layers
perforated, whereas for the cases in 4, the results would be
compared with the results of the conventional methods that used
the scenario in which the layers that are perforated must satisfy
specified constraints. Before an explanation of the results of the
four cases given above, a comparison between the performance of
a conventional GA and that of the GA modified to determine the
locations of the well will be explained.
8.1. Comparison of a conventional GA and a modified GA
To compare the performances of the conventional GA and
modified GA, a comparison test of the maximum fitness value
over time has been conducted. The conventional GA being com-
pared was taken from a GA that was made in studies for deter-
mining the locations of the wells of previous studies (Ariadji et al.,
2012). As a comparison, both of the GAs would be used in solving
case 1 above.
The results of the conventional Genetic Algorithms application
to single layer reservoir and the results of the modified GA for
multi-reservoir field that were previously conducted are shown as
follows.
The figure above clearly indicates that the modified GA has
faster convergence when seeking maximum fitness values com-
pared to that of the conventional GA. In the modified GA, the best
fitness value is already obtained at the second iteration, whereas
in the conventional GA, the fitness value can be obtained only at
the 1000th iteration or more. Noted that the fitness values of Fig. 8
(a) and (b) are not the same because Fig. 8(a) is applied only for
single layer and Fig. 8(b) is applied for multi-reservoir field, i.e., it
has summation of single layer objective function of Case 1.
In this case, only the maximum fitness values are being
compared, while we account for the fact that the conventional
GA attempted to find only one location for a well. In other words,
the conventional GA is seeking only a single optimum solution,
namely, the largest (maximum) fitness value.
The modified GA can provide faster convergence than the
conventional GA because when using the conventional GA, every
new individual that results from a crossover will definitely become
a member of the next population (next generation), whereas with
the modified GA, the fitness of the new individual will be assessed
first. If the fitness value of the new individual is better than
its parents, then the individual will be a member of the next
 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82
Fig. 8. Performance of the fitness values with respect to the number of iterations that resulted from (a) the conventional GA (Ariadji et al., 2012) and (b) the modified GA.
Table 2
Coordinate results of using the conventional trial-and-error
reservoir simulation method.
Well X Y
Well 1 19 23
Well 2 9 23
Well 3 7 28
Fig. 9. Locations of wells from using the conventional methods (trial-and-error)
(Ariadji et al., 2012).
population. In contrast, if it turns out that the new individual is
worse than its parent(s), it will subsequently not be allowed to
exist in the next population. Such a process in a modified GA
causes the newly formed population to be always superior to or at
least equal to the previous population.
8.2. Evaluation of the field X
The Field X reservoir model has 12 layers of reservoirs, with
392,812 grid blocks, and for this field, there will be three infill
wells drilled. The best three locations for the wells are those that
provide the largest oil recovery.
In the assessments of the wells, an evaluation that used a
reservoir simulator with two scenarios was conducted. The first
scenario is when all of the layers are perforated, while the second
scenario is when the perforations are conducted only in certain
layers that meet certain reservoir engineering criteria.
The GA parameters that are used to locate the wells are shown
in Table 1.
8.3. Results of the conventional trial-and-error simulation method
The previous reservoir simulation study was conducted by
using the conventional method of trial-and-error to find the 3
best locations for the production wells, with the results shown
75
Table 3
Coordinate results of using the proposed GA method with
the objective function F1.
Well X Y
Well 1 5 23
Well 2 22 26
Well 3 12 31
Fig. 10. Locations of wells using the modified GA with the objective function of F1.
in Table 2 (coordinates) and Fig. 9 (well locations on an oil
saturation map). The trial and error method was obtained by
choosing the grid block location of the infill well based on the
highest value of oil saturation, porosity and permeability, and
running it for the reservoir simulation. Then its recovery factor is
calculated by dividing the resulted cumulative oil production by
the original oil in place. This step is redoing for other location, and
so on. Of the three wells, as much as 10,634 MMSTB or a recovery
factor of 37.46% was obtained for 12 years of forecasting.
8.4. Results of the modified GA method while using the objective
function F1
By using the objective function F1, the GA provides 3 well locations,
as shown in Table 3 (coordinates) and Fig. 10 (well locations).
The three wells were then evaluated by using the reservoir
simulator software. From the simulation of the three wells above,
as much as 10.719 MMSTB of field cumulative oil production or a
recovery factor (RF) of 37.76% were obtained. The comparison of
the results of using the GA and the conventional method is shown
in Fig. 13. Even though the result is very similar to the previous
result, their production profiles are different, whereas the pro-
posed GA is better.
Based on these results, it is clear that the Modified GA with the
F1 objective function gives better results when compared to using
conventional methods.
76 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82
Fig. 11. Locations of wells from using the modified GA with the objective
function F2.
8.5. Results of the modified GA method with the objective function F2
The GA method with the objective function F2 provides 3 well
locations, as seen in Table 4 and Fig. 11.
The results of the evaluation of the reservoir simulation show
that the three wells yield oil cumulative field production that
amounts to 11.441 MMSTB or a recovery factor (RF) of 40.29%.
This objective function turns out to give better results than the
previous two methods. A comparison of three recent results is
shown in Fig. 13.
These last results are better compared to the previous results.
This comparison indicates that when pressure parameters take
part in the objective function, better results are obtained.
8.6. Results of the GA method when using the objective function F3
In the F3 objective function, in addition to the parameters of
permeability, porosity, saturation, and pressure, the grid thickness
parameter is also applied. The GA with the objective function
F3 gives the locations of the wells, as shown in Table 5 and Fig. 12.
Locating wells that are proposed by the GA with the F3
objective function is not very different from locating wells that
are proposed by the F2 objective function. A difference is found
only at the first well, which has shifted by one grid. Compared to
the previous results, at first glance, it appears that the first well is
closer to the edge of the reservoir. Usually, wells with such
locations are not at a good placement considering that the wells
that are located too close to the edge could very probably give a
small radius of drainage. However, following an evaluation by
using a reservoir simulator, it turns out that these wells provide
the same results when compared to those of previous results.
These wells produced field cumulative oil production that
amounts to 11.442 MMSTB or a recovery factor (RF) of 40.30%. A
comparison between the current results and the previous results is
shown in Fig. 13.
From Fig. 13 above, we can see that the light blue line (GA with
the F3 objective function) coincides with the black line (GA with
the F2 objective function), but the actual GA with the F3 objective
function gives very slightly better results than the GA with the F3
objective function for this field case. All of the above results are
summarized in Fig. 14 and Table 6.
8.7. Results of the conventional trial-and-error methods with
constraint criteria for selecting layers to be perforated
Unlike the other methods, in which all of the layers were
perforated, this time, the layers to be perforated must satisfy
certain constraints, such as the following:
Criterion 1: multiplication of reservoir permeability and thick-
ness, namely, C1, and
Fig. 12. Locations of the wells when using the modified GA with the objective
function F3.
Criterion 2: oil saturation, namely, C2.
For the Field X, it is determined that C1 ¼ 100 mD. m and
C2 ¼0.5. Based on the above information and criteria, the
locations of the wells as a result of the conventional trial-
and-error method provide the following Table 7 information.
Following the criteria that were previously determined, the
three wells underwent a total of 26 perforations.
Evaluation on the reservoir simulator provides information that
the above-mentioned well mechanism yielded field cumulative oil
production that amounts to 10.457 MMSTB or a recovery factor
(RF) of 36.83%.
8.8. Results of the proposed GA method with constraint criteria for
selecting the layers to be perforated
With the pre-determined criteria, this Proposed GA would
search for the 3 best well locations with the objective function
F3. This objective function is chosen for this case because it
provides the best result in the previous scenario (perforation on
all of the layers). Moreover, remembering that the constraints
involved the parameter of grid thickness, the chosen objective
function is the one that involved the thickness.
With all of the information and criteria mentioned above, the
proposed GA provides the well locations and combination of layers
perforated, as seen in Table 8.
The combination of the locations of the wells and the perfo-
rated layers above provides field cumulative oil production
that amounts to 11.267 MMSTB or a recovery factor (RF) of 39.69%.
This result is better when compared with the results of the
conventional methods, especially when the number of perfora-
tions is lower. Thus, the proposed GA is more economical when
compared with the results of the conventional methods with more
perforations.
Some interesting findings of this proposed GA is that the
resulting well locations are exactly the same as those in the
previous Modified GA with the F2 Objective Function (all layers
perforated). Looking at the comparison of the results in the
previous scenario (perforated on all layers), namely, the result of
the Modified GA with the F2 objective function and the GA with
the F3 objective function, the GA provides similar locations for the
proposed wells. The only difference is in the location of the first
well. In the first well, the GA with the F2 objective function is
located at (22,26), and the GA with the F3 objective function is
located at (23,25), which constitutes a shift of only one grid. The
oil recovery is almost the same.
Table 9 and Figs. 15 and 16 summarize a comparison of the
resulting well locations for the scenarios of all of the layers to be
perforated and the selected perforated layers.
 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82 77

Fig. 13. Forecasting a comparison of the field cumulative oil production that results from the conventional GA method and the modified GA with the objective functions F1,
F2, and F3. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Table 4 Table 6
Results of using the proposed GA method with the Results for the scenario with all layers perforateda.
objective function F2.
Methods Well 1 Well 2 Well 3 Prod. cumulative Recovery factor
Well X Y (MMSTB) (%)

Well 1 22 26 Conventional (19,23) (9,23) (7,28) 10.634 37.46

Well 2 7 28 GA with F1 (22,26) (5,23) (12,31) 10.719 37.76
Well 3 12 31 GA with F2 (22,26) (7,28) (12,31) 11.441 40.29
GA with F3 (23,25) (7,28) (12,31) 11.442 40.30

a
Perforation performed when all of the layers are commingled at the
same time.

Table 5
Results of using the modified GA method with the Table 7
objective function F3. Results of selected perforated layers using the conventional method.

Well X Y Well X Y Layers perforated Number of perforations

Well 1 23 25 Well 1 19 23 1,2,3,7,8,9,11 7

Well 2 7 28 Well 2 9 23 1,2,3,7,8,9,10,11,12 9
Well 3 12 31 Well 3 7 28 1,2,3,6,7,8,9,10,11,12 10

Table 8
Results of the well location and their selected perforated layers that were obtained
from applying the perforation criteria of the proposed GA method with the
objective function F3.

Well X Y Layers perforated Number of perforations

Well 1 22 26 1,2,3,4,5,6 6
Well 2 7 28 1,2,3,6,7,8,9,10,11,12 10
Well 3 12 31 1,2,3,4,5,6 6

8.9. Inclusion of a drainage radius, existing wells, existence of faults

Fig. 14. The diagram of the results summary of the scenarios with all layers
perforated*.
The above results illustrate that the GA consistently provides a
better combination of the well locations according to the specified
criteria for the layers in this field. This finding indicates that the
proposed GA method is more effective (better oil recovery) and
efficient (faster) than the conventional GA.
and the relevant software development
The above Genetic Algorithm method, which is formulated
specifically for reservoir engineering well placement plans, has
been further extended to consider the drainage radius, existing
wells and existence of faults, all at the same time. Furthermore, a
software application has been developed that incorporates these
concerns, as depicted by the following flowchart in Fig. 17.
78 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82

8.9.1. Drainage radius data input 8.9.2. Faults

The drainage radius, which is one of the important key parameters Pertaining to migration pathways, faults have the role of
in oil-field monitoring, acts as a necessity in the determination of the leading to traps or barriers to flow. A well that is located near
well locations. Knowing the drainage radius helps engineers to faults could result in the lessening or slackening of its depletion.
optimize the locations of new wells to be drilled in the field. The Its drainage area might be less than the actual drainage area that
estimation of the drainage radius can be provided by well test analysis. the well can produce. Therefore, the existence of faults should not
Otherwise, the developed software offers a selection as to whether the be ignored when determining the location of wells.
drainage radius is determined by well test analysis or by an assump- Upon the completion of searching for the optimum wells, the
tion of the well spacing. The assumption that is used in the software is software reads the inputted fault trajectory data. Afterward, the
that the drainage area is circular; hence, the drainage radius can be software will mark the grids that are located near to the faults
calculated by reversing the equation of the circle area. depending on the drainage radius. The drainage radius that is
entered by the user is converted into numbers on the grid. The
illustration of this process is represented in Fig. 18, in which the
darkened grids are the marked grids or the grids that are located
Table 9 near faults that have a drainage radius that is equivalent to one
Results that show a summary of oil cumulative production scenarios. grid. Normally, if no fault exists, then the ranking process would
depend on the fitness value of each grid, and the prioritization
Methods Well 1 Well 2 Well 3 Prod. RF No. of
would be purely derived from the fitness value. However, if faults
cumulative perforations
(MMSTB) (%)
do exist, then the marked grids are not prioritized when traversing
the ranking process. The grids would not be deleted from the
Conventional (19,23) (9,23) (7,28) 10.457 36.83 26 possible solutions, but their ranks would be lowered.
GA with F3 (23,25) (7,28) (12,31) 11.442 39.68 26
for all layers
perforated
GA with F3-selected (22,26) (7,28) (12,31) 11.267 39.69 22
8.9.3. Existing wells
perforated
layers Consideration of the previous production wells in a field should
not be ignored. If previous wells actually exist in a field and the
software does not account for them, then a very large error can
occur in the determination of the infill wells. The drainage of the
new wells can overlap the drainage of the previous wells in such a
way that their productivity will decrease. Thus, the developed
software provides the option of whether previous wells do exist or
not. If there are existing wells, then the users are allowed to input
their locations, their drainage radius and/or cumulative production
data of the existing wells. In the condition in which the user
entered both the drainage radius and the cumulative production
data, the drainage radius data would be preferred and the software
will disregard the cumulative production data. The equation below
would be used when the user enters only the cumulative produc-
tion data to estimate the drainage radius.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
N p  Bo
rad ¼ ð6Þ
Fig. 15. The diagram that shows a summary of oil cumulative production scenarios. π  h  ϕ  So

Fig. 16. Forecasting comparison of the cumulative oil production performance for scenarios using the conventional method and the modified GA with F2 (for all layers
perforated and for selected perforated layers).
 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82
Fig. 17. Flowchart for developing the software application.
Fig. 18. Illustration of the marked grids as a consequence of their locations near
faults.
79
where
rad ¼drainage radius of well (ft),
Np ¼cumulative oil production (MMSTB)
Bo ¼oil formation volume factor (bbl/STB),
h¼reservoir thickness (ft),
ϕ ¼porosity,
So ¼ oil saturation.
Similar to the treatment of the grids that are located near faults,
the grids that are located near the existing wells would also be
marked by the software, depending on the drainage radius of each
well. The marked grids are less prioritized in the process of ranking;
thus, their ranks would be lowered and other grids that are neither
near the fault nor near existing wells could arise. Fig. 19 shows the
illustration of the marked grids that are located near the faults.
8.9.4. Multiple layers
The majority of the fields are currently produced by multi-
reservoirs. The layers vary in their rock properties, such as perme-
ability, porosity and saturation, which cause each layer to have
different productivity. The location determination of the multi-
reservoir wells is more complex than the location determination of
single reservoir wells. Consequently, the best location for a single
layer reservoir well would not be the same as for a multi-reservoir
well. Thus, the developed software is definitely required to opti-
mize the well placements of multi-reservoir fields.
Users are given the option of whether to consider all of the
layers or to consider only the layers that surpass the criterion for
selective perforation. If the selective perforation option is selected,
then the users must provide the data of the minimum kh and
the hydrocarbon saturation. Furthermore, the developed software
allows the users to determine the maximum number of layers that
are to be perforated. For substantiation, the fitness values of a well
that has 50 perforated layers and a well that has only 10 perforated
layers are not comparable. The well with 50 perforated layers
could have a higher fitness value, but in regard to only perforating
the top 10 layers, the conditions would be different. Therefore, by
entering the maximum number of layers into the software, the
fitness value of each well is calculated from the best layers with
the same number of layers as that entered by the user. This feature
would be favorable in the case when a company has a budget that
is sufficient for perforating only a certain number of layers.
8.9.5. Validation
To verify the pertinence of the developed software, a validation
phase is required. The validation phase utilizes a reservoir
Fig. 19. Illustration of the marked grids as a consequence of their locations being
near to existing wells.
80 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82

simulator and compares the oil recovery and recovery factor of the Table 11
3 proposed wells with the wells that are obtained by using the Well placements from using the previous Genetic Algorithm study.
previous study and the conventional method. Because the con-
Well X Y Perforated layers
ventional method and the previous study utilized only 12 layers, to
make the results comparable, this study also searched for 12 Well 1 23 25 5,6,7,8,9,10,20,21,22,23,24,25
optimum layers to be perforated. The parameters that are used Well 2 7 28 5,6,7,8,9,10,20,21,22,23,24,25
in the validation phase can be seen in Table 1. The simulation was Well 3 12 31 5,6,7,8,9,10,20,21,22,23,24,25

run for 11 years, with a minimum bottom hole pressure of

1200 psia and a tubing size of 7.5 in. The drainage radius was set
to 250 m or 820.21 ft because it is the rule of thumb drainage
radius for the proven reserves. To ensure economic limits, the
maximum water cut allowed was 98%, and the minimum flow rate
for each well was 5 STB/day.

8.10. Software validation

We used the resulting well locations from the conventional

trial-and-error method, as shown in Table 10.
A previous attempt of locating the optimum well placements
has been performed, and the result is summarized in Table 11. The Fig. 20. Well placements using the developed software.
previous study used only the Genetic Algorithm in determining
the (x,y) location and not the layers that are going to be perforated.
Table 12
Therefore, it can be seen that the conventional and previous study Oil recovery comparison of the trial-and-error technique and the proposed method.
have the same perforated zones. These zones are determined by
the estimation and intuition of the reservoir engineer. As a human, Method Cumulative oil production Recovery factor
the reservoir engineer also tends to simplify things and perceives (STB) (%)

that the perforated layers of one well would be the same as the
others. As a result, a human error could occur in the determina-
tion; thus, the risk of obtaining the wrong solution might increase.
According to the developed software, the three most suitable
wells for the field, which can be seen in Fig. 20, have been chosen.
The perforated layers are now also determined, to achieve an
improvement in the oil productivity.
A significant improvement in the oil recovery when using the
developed software is shown in Table 12. An increase of 3.32%
for the recovery factor compared to the trial-and-error method
occurs. Furthermore, compared to the previous study, which did
not consider the existence of faults, the multi-layer reservoir
selection and the drainage radius, the new proposed wells show
a satisfactory result with a recovery factor enhancement of 0.49%.
In other words, the utilization of the developed software can
provide an extra 142 MSTB of oil compared to the previous study,
or an extra 944 MSTB of oil compared to the conventional trial-
and-error method. This increase in the oil recovery will surely lead
an increase in the economic benefits.
Trial-and-error 10,639,866 37.47
Previous study 11,442,057 40.30
Proposed Software 11,584,528 40.79
maximize the length of plateau time. This paper addresses this
optimization.
For the above reason, we have expanded the GA method to a
horizontal well placement by using basic reservoir rock properties
to develop an objective function and by combining the GA with the
Artificial Neural Network (ANN) method to maximize gas produc-
tion. Again, this method was also tested by a real field, namely, the
Z Field, with a data set as described in the following. The Z gas
field has two productive zones, and a horizontal well would
be located along those two productive zones. The reservoir fluid
is identified as dry gas reservoir (which has a very high methane
component), and the reservoir has good porosities and fair
permeabilities. Thus, fluid flowing into the horizontal wellbore is
considered to be a single phase dry gas, and no condensation
occurs.

8.11. A combination of a genetic algorithm and an artificial neural

network for horizontal well placement 8.12. A brief description of the Z field reservoir model

A gas field development is different from an oil field develop- The Z Field is an offshore gas field at an early stage of develop-
ment, i.e., the gas demand from a buyer is set prior to gas ment. Gas bearing formation of this field belongs to formations of
production to be delivered. A mount of the gas demand is specified late Miocene to late Pliocene ages. The formations consist of
through a gas flow rate that is planned to be held for a certain planktonic foraminifera of wackestone to grainstone facies depos-
period of producing time, namely, gas plateau rate, and, the length ited in outer-neritic to upper-bathyal setting. The gas trap in Z field
of producing time that is called by plateau time. Thus, having is related to the combination of anticlinal and faults structure
set the gas flow rate, optimizing the gas field production is to along with stratigraphic traps as the result of deep channeling in
the shaly formation.
Table 10
The platform location limits the well drainage, and thus, a
Well placements using the conventional trial-and-error method.
horizontal well is probably an unavoidable development scenario.
Well X Y Perforated layers Fig. 21 shows the condition of the reservoir grid cells.
As a consequence of having poor rock properties towards the
Well 1 19 23 5,6,7,8,9,10,20,21,22,23,24,25 east direction, a horizontal well placement would be directed to
Well 2 9 23 5,6,7,8,9,10,20,21,22,23,24,25
Well 3 7 28 5,6,7,8,9,10,20,21,22,23,24,25
the left side, and hence, would accurately determine which
direction is the goal of this horizontal well placement.
 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82 81

direction, is (Ariadji et al., 2012)

 
u v
α ¼ 2π  cos  1 ð9Þ
‖u‖  ‖v‖
N
where u ¼ ðx0  x⋯y0 yÞ and v ¼ð0⋯1Þ.

8.14. Neural network architecture

This study uses the generalized method with a back propaga-

tion network (BPN) that is a multi-layered network, in which
information flows from the input to the output through at least
one hidden/middle layer. Each layer contains neurons that are
connected to all of the neurons in the neighboring layers. The
connections have numerical values (weights) associated with
them. The transfer function between input layer and hidden layer
is sigmoid function, and the transfer function between hidden
layer and output layer is linear function. The optimum transfer
Fig. 21. The Z Field grid cells. function between input and output layers is determined by trial
and error. During the training phase, the weights are adjusted
according to the generalized delta rule. Training is completed
when the network is able to predict the given output.
Collectedφ, k, S Generated HW direction Obtain the optimum
data from a objective function for the direction: Run the
reservoir model proposed GA proposed GA 8.15. Results and analysis of the horizontal well case

The GA simulation obtained an optimum direction of a hor-

izontal well from a known well location at a platform, in the
Validate the Obtain the optimum Collected plateau time,
results with L: Run Artificial Σφ,Σk,ΣS and Σφ, k, S data direction of 3481. Following the optimum direction, we generated
reservoir Neural Network from a reservoir model
the data of the horizontal well lengths (L) with respect to the
plateau time, with respect to the summation of each grid value for
Fig. 22. Flow chart of the horizontal well optimization using a combination of the the permeability (Σk), saturation (ΣS), and porosity (Σϕ), and the
GA and ANN methods.
multiplication of the permeability, saturation and porosity (ΣkSϕ)
for the corresponding direction. In this study, we use only three
Fig. 22 is an approach in which the horizontal direction is data for L (L¼ 50 m, 350 m and 650 m) to obtain accurate and
determined first by the GA method and the length is determined acceptable results when using the Artificial Neural Network (ANN).
by the ANN method, while using the several reservoir simulation Table 13 shows the input data for the ANN that was obtained from
results as the training data. All of the results in this study then will the reservoir model generated by ECLIPSE.
be validated by using commercial reservoir simulator, i.e. ECLIPSE The ANN yields the optimum horizontal well length of 600 m,
which is used most commonly by reservoir engineers. which results in 6.75 years of plateau time with an average error of

Table 13
Input data for the ANN (obtained from the commercial reservoir simulator).

8.13. The objective function for the direction of the horizontal L ∑k ∑ϕ ∑S ∑kϕS Plateau time
well placement (m) (yr)

The objective function at (x,y) is a simple formula compared 50 60.27 0.64 1.78 17.31 3.64
350 247.79 2.69 7.09 74.01 6.69
with the previous formulas and is written below, as follows
650 431.36 4.75 12.19 127.4 6.50
(Ariadji et al., 2012):

f ðx; yÞ ¼ ∅ðx; yÞ  Sðx; yÞ  Kðx; yÞ ð7Þ

Note that f has no physical interpretation, but it is sufficient to Table 14

Comparison of the proposed method and the reservoir simulation results.
represent a model for optimizing the production rate. Next, let
(xo,yo) be the horizontal well location. Let (xi,yi) represent the grid L Plateau time Difference
by the horizontal, which is located between (xo,yo) and (x,y). To (m) (yr) (%)
find the optimum direction of the horizontal well, the objective
function would be (Ariadji et al., 2012) GA–ANN Simulator

100 6.52 4.93 32.15

Fððx0 ; y0 Þ; ðx; yÞÞ ¼ ∑ f ðxi ; yi Þ ð8Þ 150 6.54 5.63 16.13
i
200 6.58 6.05 8.74
250 6.64 6.29 5.40
where (x,y) in Eq. (8) represents the horizontal well tip point. 300 6.69 6.53 2.51
Hence, we can obtain the optimum direction by using the 400 6.74 6.82 1.20
(xo,yo) and (x,y) coordinates, as mathematically explained below. 450 6.74 6.92 2.52
Suppose that vector u is on a vector that is constructed by the 500 6.75 6.98 3.39
550 6.75 6.98 3.38
horizontal well, and vector v is a north vector. Thus, the optimum 600 6.75 6.82 1.08
direction of the horizontal well (α), in the counter clockwise
82 T. Ariadji et al. / Journal of Petroleum Science and Engineering 122 (2014) 69–82

8 9. Conclusion
7
The following comments conclude the above study:
Plateau time, years

6 (1) A proposed method that uses a genetic algorithm-based

5 software application has been developed to optimize vertical well
Reservoir placements, and it is a robust optimization method because the
4 Simulator objective function involves only basic reservoir engineering prop-
3 GA-ANN erties, i.e., porosity, permeability, saturation, and pressure maps. In
Results other words, this proposed method does not need to run a
2
reservoir simulation model, which is often the largest challenge
1 of an optimization method. However, an application of the devel-
0 oped software can offer better results and a higher recovery
0 200 400 600 800 compared with the conventional trial-and-error methods. (2) In
Length, m the case of horizontal well optimization, the proposed method
requires a combination of the Genetic Algorithm method and the
Fig. 23. The comparison between the proposed method and the reservoir simula-
Artificial Neural Network approach for optimizing the direction
tion results.
and horizontal segment length, respectively. This proposed
approach could offer a faster and acceptable result when deter-
mining the optimum of the horizontal well direction and the
Table 15 horizontal segment length.
Comparison of the commercial reservoir simulator runs.

Method α L Plateau Number References

time of runs
(deg) (m) (yr)
Ariadji, T., Soekarno, P., Sidarto, K.A., Soewono, E., Riza, L.S., Kenny, D., 2012.
Optimization of vertical well placement for oil field development based on basic
The conventional 315 696 7.17 165
reservoir rock properties using genetic algorithm. ITB J. Sci. 44 (1), 106–126.
trial-and-error Ariadji, T., Aziz, P.A., Soewono, E., Syifa, A.A., Riza, L.S., Sidarto, K.A., Sukarno, P.,
The proposed 348 600 6.75 3 2012. A robust method for determining the optimum horizontal well direction
method and length for a petroleum field development using genetic algorithm In: AIP
Conference Proceedings, vol. ICREM5.
Bittencourt, A.C., Horne, R.N., 1997. Reservoir development and design optimiza-
7.65%. Table 14 and Fig. 23 show the comparison between the tion. In: Proceedings of the SPE Annual Technical Conference and Exhibition,
Paper SPE 38895. Texas.
proposed and reservoir simulation results.
Dianati, M., Song, I., Treiber, M., 2002. An Introduction to Genetic Algorithms and
Our previous study using a trial-and-error procedure yields the Evolution Strategies. University of Waterloo, Ontario, Canada.
optimum direction and horizontal well length of 3151 and 696 m, Mitchel, M., 1996. An Introduction to Genetic Algorithms (Section 1.6. A Simple
respectively, which results in 7.17 years of plateau time. We can Genetic Algorithm). MIT Press, pp. 8–9.
Montes, G., Bartolome, P., Udias, A.L., 2001. The use of genetic algorithms in well
see that the resulting plateau time between the proposed method placement optimization. In: Proceedings of SPE Latin American and Caribbean
and the conventional trial-and-error method is quite small. How- Petroleum Engineering Conference, Paper SPE 69439. Argentina.
ever, the proposed method simulation yields a shorter horizontal Ozdogan, U., Horne, R.N., 2004. Optimization of well placement with a history
matching approach. In: Proceedings of the SPE Annual Technical Conference
segment length, which means that the proposed method gives and Exhibition, Paper SPE 90091. Texas.
better results economically. Moreover, to obtain these results, the Pham, D.T., Karaboga, D., 2000. Intelligent Optimization Techniques: Genetic
proposed method required much less effort from the conventional Algorithms, Tabu Search, Simulated Annealing and Neural Networks. Springer,
London.
reservoir simulation. Thus, the proposed method could be an
Romero, C.E., Carter, J.N., Gringarten, A.C., Zimmerman, R.W., 2000. Improve
efficient and effective solution for obtaining an optimum horizon- reservoir characterization through evolutionary computation. In: Proceedings
tal well placement and length (Table 15). of the SPE Annual Technical Conference and Exhibition, Paper SPE 62942. Texas.