Sie sind auf Seite 1von 18

SPE 113525

Optimization of Well Placement Using Evolutionary Algorithms


D.Y. Ding, SPE, IFP

Copyright 2008, Society of Petroleum Engineers


This paper was prepared for presentation at the 2008 SPE Europec/EAGE Annual Conference and Exhibition held in Rome, Italy, 912 June 2008.
This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents of the paper have not been
reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect any position of the Society of Petroleum Engineers, its
officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written consent of the Society of Petroleum Engineers is prohibited. Permission to
reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
Non-conventional wells allow to increase considerably hydrocarbon recovery. By considering the high drilling cost and the
potential earn in well productivity, optimum implementation of non-conventional wells is an important issue in petroleum
industry.
Considering large number of parameters involved in well placement and high reservoir heterogeneities, stochastic
methods such as evolutionary algorithms are the most efficient approaches for optimization. In this paper, we present the
CMAES (Covariance Matrix Adaptation Evolution Strategy) method, which is based on evolutionary approach and has
been considered as one of the best stochastic optimization method for non-linear problem. The application of CMAES to the
problem of well placement optimization is presented. CMAES is an alternative approach for the well placement, and it gives
comparable results with respect to the genetic algorithm.
Although the evolutionary methods such as CMAES or genetic algorithm are suitable for the modelling of well
placement, their efficiency depends on various parameters involved in the model. In this paper, the impacts of model
parameters in the optimization of well placement are also discussed.

Introduction
Non-conventional wells allow to increase considerably hydrocarbon recovery. The productivity of this kind of wells depends
on various factors, such as well configurations (length, orientation, number of branches, ...), reservoir heterogeneity, reservoir
fluid (water, oil, gas), etc. All these make the optimum well implementation difficult. By considering the high drilling cost
and the potential earn in well productivity, optimum implementation of wells is extremely important in field development
plan, because it can significantly affect the project's Net Present Value (NPV). The capital investment required to drill a well,
especially a non-conventional well, is usually very high.
Selection of optimum well trajectory for non-conventional wells is extremely challenging. In a field development context,
we need to find the optimum number of injection and production wells to be drilled, their locations, trajectories as well as
their types such as number of branches, optimum junction points, length and orientation of branches. A large number of
parameters are involved. Nevertheless, optimization of well placement can be achieved by numerical simulations.
Considering large number of parameters involved and high reservoir heterogeneities, stochastic methods such as evolutionary
algorithms seem to be the most efficient methods for optimization.
Evolutionary algorithms have been developed since the 1960s. Nowadays, there are mainly four branches: genetic
algorithm (Holland [1]), evolution strategy (Rechenberg [2]), evolutionary programming (Fogel et al. [3]) and genetic
programming (Koza [4]). The genetic algorithm has already been applied in petroleum engineering to optimise well
placement for vertical wells (Bittencourt and Horne [5] et Gyagler et al. [6]) and also for non-conventional wells to
determine well trajectory, length, branches etc. (Yeten et al. [7], Artus et al. [8]). In those applications, optimization
parameters are discretized to form chromosomes. The efficiency of the genetic algorithm depends on the manner of
discretization.

SPE 113525

In this paper, we will present an application of the approach CMAES (Covariance Matrix Adaptation Evolution
Strategy), a variant of evolution strategy, to the optimization of well placement. The evolution strategy (ES) is an
optimization technique based on ideas of adaptation and evolution. It was created in the 1960s and 70s by Ingo Rechenberg et
al. [2]. It has been applied to various optimization problems, including continuous, discrete or multi-objective function
problems. The adaptation technique (Hansen and Ostermeier [9], [10], Hansen [11]) has been developed to adapt the
covariance matrix of the multivariate normal mutation distribution in the evolution strategy (ES). This method has been
studied by many authors (Michalewicz [12], Eiben and Smith [13], Auger [14]).
In this paper, the approach CMAES is compared with the genetic algorithm through an example, using simple attributes
as objective function. Comparable results are obtained with both methods. Although the evolutionary methods such as
CMAES or genetic algorithm are suitable for the modelling of well placement, the efficiency of these methods depends on
various parameters involved in the model. For example, the efficiency of CMAES depends on the step size or the learning
rate, while the efficiency of the genetic algorithm depends on the discretization of the parameters or the size of the binary
chromosome. Sensitivities are studied to evaluate the impacts of model parameters in the optimization of well placement.
A new approach based on calculation between simple attributes and NPV is proposed to optimise well placement in terms
of NPV. This approach allows to reduce considerably number of reservoir simulations as well as CPU time.

Approach CMAES
In this section, we give a brief presentation of the basic idea of CMAES. Detailed descriptions can be found in Hansen [11].
The method CMAES is a stochastic method combining two different notions with ES for mutation and selection etc. and
CMA for adaptation of covariance matrix. Evolution strategies use natural problem-dependent representations, and primarily
mutation and selection as search operators, which are applied in a loop. An iteration of the loop is called a generation. The
sequence of generations is continued until a termination criterion is met. The mutation distribution is used to sample new
candidate solutions. The covariance matrix adaptation (CMA) is a method to adapt the covariance matrix of the multivariate
normal mutation distribution in the evolution strategy. It describes the dependencies between the variables in the distribution.
Let x be a point (individual) in the parameter space and f(x) be the objective function at the point x, we want to find the
optimum parameter x so that f(x) attend the maximum (or minimum). We will search the point x in a stochastic framework. It
is assumed that the population of search points is generated by sampling a multivariate normal distribution N(m, C) with
mean m R n and covariance matrix C R nn .
Let the population size in a generation (number of points in an iteration) and g the generation number, the new search
points in the generation g+1 is sampled from the following formula:
for k = 1, ... ,
xk( g +1) ~ N (m ( g ) , ( ( g ) ) 2 C ( g ) )
where
~ denotes the same distribution on the left and right side,

xk( g +1) R n is the k-th offspring (search point) from generation g+1,
m ( g ) R n is the mean value of the search distribution at generation g,

( g ) R+ corresponds to the step size at generation g,


C ( g ) R nn is the covariance matrix at generation g. Up to a constant factor, it is the covariance matrix of the
search distribution.

In the optimization algorithm, the key problem is to determine the distribution function for the next generation g+1, that
is, to estimate the mean m ( g +1) , the covariance C ( g +1) and the step size ( g +1) for the generation g+1.
In the generation g+1, we have generated points x1( g +1) , ..., x( g +1) . The new mean of the search distribution is

calculated using a weighted average of selected best points according their values in the objective function:

m( g +1) =

i =1

where

wi xi(:g +1)

SPE 113525

i =1

wi = 1,

w1 w 2 ... w > 0

with wi the weight, , xi(:g +1) is the i-th point among x1( g +1) , ..., x( g +1) with f ( x1(:g +1) ) f ( x2( :g+1) ) ... f ( x( g:+1) )
and f is the objective function to be maximized (or minimised in a minimisation procedure). Default weights proposed by
Hansen [11] are used in the optimization searches.
The new covariance matrix for the search distribution is calculated by:

T
C ( g +1) = wi xi(:g+1) m ( g ) xi(:g+1) m ( g )
i =1

)(

The above covariance matrix is estimated from selected points. To assure C( g +1) is a reliable estimator, the population
size should be large. But in most cases, we don't have sufficient points. So, variance techniques are developed for the
adaptation of the covariance matrix. The adaptation of covariance matrix is a key point of CMAES, which is briefly described
in Appendix.

Evolutionary methods for the optimization of well placement


We give, in the following, a brief review of the genetic algorithm which has already been used for well optimization (Yeten
et al. [7]), and show how to adapt the CMAES algorithm to the well placement.

Parametrization
Wells are parametrized as proposed by Yeten et al. [7]. A well is represented by one or several straight line segments. The
main drain is described by 6 parameters: the coordinates (xh, yh, zh) of the heel, an angle xy (angle between its projection in xy
plan and the x axis), the horizontal length Lxy (well projected to xy plan) and the vertical length Lz (well projected to z
direction). In the presence of branches, the junction position is determined by the its distance Lb to the well heel, and the
branch trajectory is again described by using xy, Lxy and Lz.
Genetic algorithm
The genetic algorithm (Yeten et al. [7]) needs optimization parameters in discretized forms. For any parameter B to be
optimised, we define a lower bound Bmin, an upper bound Bmax, and a discretization step B . Possible values of the parameter
B are given by
B
Bmin
with the integer i 0, max
.
Bi = Bmin + i B
B
Once parameters are discretized, we can use binary code to represent them to form a chromosome. Smaller the
discretization step B is, longer the binary code is. Operations such as crossover, mutation and elitism are used in the
genetic algorithm. The stopping criteria is the maximum number of evaluations or the following constraint:

f ( M g) f ( M g N )
f (M g N )

<

where f(Mg) is the objective function for the best point in the generation g, N and are given data.
CMAES for well placement optimization
Same optimization parameters xh, yh, zh, xy, Lxy, Lz etc. as described above are used in the CMAES algorithm. Different from
the genetic algorithm approach, CMAES uses continuous parameters. The code developed by Hansen [11] is used in our
optimization procedure. The stopping criteria is the maximum number of evaluations for the objective function.

The parameters are constrained within Bmin and Bmax, while the CMAES algorithm searches the optimum point in Rn. To
force the search inside the defined domain, penalisations for the objective function are used when necessary. In particular, if
the lengths Lxy or Lz are too long beyond the constraint maximum lengths, penalisations are introduced in the objective
function.

SPE 113525

Objective function
The main objective of this paper is to evaluate the feasibility of CMAES in the well placement optimization and to compare
CMAES and the genetic algorithm. However, if the NPV is used as the objective function, CPU time might be too long
because several thousands of reservoir simulations are needed for NPV calculation in an optimization run. Nevertheless, we
can use other simple attributes instead of NPV to compare the efficiency of different optimization algorithms. In this paper,
two attributes, the numerical productivity index (PInum) and the field productivity index (PIfield), are considered as the
objective function in the comparison.
Numerical productivity index
The numerical productivity index (PInum) is an attribute defined on the wellblocks. Peaceman's formula is used for the
PInum calculation on a wellblock. The well flow rate on a wellblock i is proportional to PInumi on the same block:

Qi = PInumi Pi
where P is the difference between the wellblock pressure and the wellbore pressure. The attribute PInum is the sum of
well productivity index on all well blocks:
PInum =
PInumi
i

Field productivity index


Field productivity index (PIfield) is defined as follows:

PIfield =

Q
Pav Pw

where Q is the total well flow rate, Pav is the average pressure in the drainage area and Pw is the wellbore pressure. PIfield
can be obtained from well test simulation/interpretation, and it can also be approximated by using single phase flow with only
one large time step in the simulation. Calculation of PIfield using one time step is very fast.

Numerical example
Considering a reservoir of size 3000x3000x250 m3, this reservoir is discretized by 30 blocks in x direction, 30 blocks in y
direction and 5 layers in z direction. The permeability field, which is very heterogeneous, is shown in Figure 1. A well with
only main drain will be implemented in this field. Both genetic algorithm and CMAES are used to search the optimum
position of this well, according to the objective function which can be either PInum or PIfield.

In the genetic algorithm, the maximum number of evaluation is 5000, and the stopping criteria is N = 50 and = 0.1. The
ranges of parameter variation and the steps of parameter discretization are given as follows:
- well heel coordinates xh and yh vary between 0 and 3000m with discretized step 97m;
- well heel coordinate zh varies between 0 and 250m with discretized step 25m;
- angle is discretized with step 30;
- well length projected in xy plan Lxy varies between 0 and 1000m with discretized step 32.26m;
- well length in the vertical direction Lz varies between 0 and 250m with discretized step 35m.
For the CMAES, the maximum number of evaluations is also 5000. CMAES searches the parameters xh, yh, zh, , Lxy and
Lz in Rn with constraint 0 Lxy 1000m and 0 Lz 250m . Comparing the reservoir height (250m) and the maximum well
vertical length Lz authorized (250m), we can find, in many cases, a part of the well outside the reservoir. To avoid these
cases, a penalisation term is added in the objective function to force the vertical coordinates of well extremities inside the
reservoir region.
First, PInum is used as objective function. 10 runs are performed with genetic algorithm and the results are shown in
Table 1. The first column is the maximum objective function, and the second column corresponds to the number of
evaluations to get to this maximum value. A series of runs are also performed with CMAES for various population size ,
and the results are shown in Table 2. The default population size for CMAES in this example is = 9. It is found that the

SPE 113525

results are not very good when the population size is small. However, when the population size becomes larger, significant
improvements are obtained. Although the objective function can reach to high value above 64000 with small population size,
results are not homogeneous in different runs and sometimes local maximums are found. For large population size, results
are much more homogeneous. If we perform only one optimization run, which is what we can afford in most real cases, using
large population size seems more convenient.
Although CMAES provides generally higher values in objective function than the genetic algorithm, it needs usually
more evaluations. It seems that CMAES reaches its maximum value very late. However, if we analyse the evolution of the
objective function in CMAES (Figure 2), we find that relatively high values can be obtained quite early. But, as it is a
continuous approach, the objective function can be improved in a very small gain with more evaluations in objective
function.
The best well positions projected in xy plan are shown in Figure 3 for the 10 runs in genetic algorithm and in Figure 4
for the runs in CMAES with population size 9 and 60. In these figures, the permeability in the background corresponds to the
layer 3. It is found that in most cases, all algorithms search in the same region, although it happens to search in a different
region for CMAES with population size = 9. As discretized formulation is used in genetic algorithm, possible well
configurations are very limited. However, CMAES is a continuous approach, well configurations are more various.
Now, we use the PIfield as the objective function to compare the optimization algorithms. Simulation results of genetic
algorithm and CMAES are summarised in Table 3 and Table 4. Like previous comparison, CMAES gives generally higher
objective function than the genetic algorithm, but it needs more evaluations.
The best well positions for PIfield are shown in Figure 5 for genetic algorithm and Figure 6 for CMAES with population
size 9 and 60. Again, in most cases, all algorithms search in the same regions. But this region is slightly different from that
with PInum as objective function.
In this example, almost all algorithms arrive to final searches in the same region. This can give qualitative information
about the region to be drilled. Qualitative information is helpful to make decision on drilling area. However, to quantify the
best well trajectory, different runs give quite different results. Nevertheless, the CMAES can potentially provide higher
objective function. If more evaluations are allowed, we believe that almost all runs would converge to the same position, as
the search has already been localized in the same area. This example shows that CMAES is suitable for well placement
optimization, and it gives comparable results versus the genetic algorithm. But the population size is a sensible parameter in
the optimization procedure.

Sensitivity study
The efficiency of evolutionary methods depends also on model parameters. For example, the efficiency of genetic algorithm
depends on how parameters are discretized. CMAES depends on the step size . Using different model parameters might
change the searching results. The impacts of some parameters on searching optimal well position are studied in this section.
In this sensibility study, PInum is used as objective function, and we limit the maximum number of evaluation to 500.
For the genetic algorithm, we vary the discretization of the well head position xh and yh. Noting xh 0 = 97m and
yh 0 = 97m the initial discretization steps for the well heel coordinates, we denote GA2n the genetic algorithm by
multiplying the initial steps by a factor of 2n. That is, the discretization steps in the genetic algorithm GA2n are

xh = 2 n xh 0 and y h = 2 n y h0 .
For each algorithm GA2n with discretization steps 2 n xh0 and 2 n yh 0 , 10 runs were performed. Figure 7 shows
simulation results with different discretization steps. The orange line corresponds to the maximum value obtained for each
discretization and the red line corresponds to the average value. When the discretization steps are too large (large n value),
we have a small set of possible configurations (see n = 4 in Figure 8). Consequently, both the maximum and average values
in objective function are not very high. When the step is too small (for example n = -13), there are too many possible
configurations. The global search has not identified the good searching region by limiting the maximum evaluation to 500.
Consequently, the searching results are very dispersed (Figure 8). Although the best result gets high objective function
(63393), the average value is not very high (42075). It seems that neither large nor small discretization steps is adapted for
the well optimization. The intermediate steps, which correspond to about 1/10 of gridblock sizes (n = -3 or -4), seem the most
suitable. The best well positions with the intermediate step (n = -3) are also shown in Figure 8. The searches are focused in
the same region with various well configurations.

SPE 113525

The CMAES optimizations with default model parameters are performed for population size = 9, 20, 40 and 60 with the
maximum number of evaluations fixed to 500. The results are shown in Figure 9 with average values represented by the red
curves. The thin red curve with square symbol corresponds to the average for each population size and the thick red curve
with circle symbol corresponds to the global average for fixed model parameters. Using default parameter in this example,
average values are quite low. However, if we change model parameters by multiplying the learning rate Ccov by a factor of 2,
the results are generally improved. If we again reduced the step from 0.5 (default value) to 0.3, results are still generally
improved (Figure 9). These results show that the efficiency of CMAES depends also on model parameters. Figure 10 shows
well positions with = 20 using default model parameters and with = 20 by multiplying Ccov by a factor of 2. In the first
case (default parameters), well positions are very dispersed and searching regions do not converge, while in the second case
(Ccov multiplied by 2), almost all wells converge to the same region. So, results in the second case are better than those in the
first one. These results show that model parameters might have a great impact on the optimization. CMAES can be
potentially improved with good choices of model parameters. However, determining best model parameters in well
placement needs supplementary studies.

Optimization of well placement in terms of NPV


It might be very expensive in CPU time to optimise well positions using directly NPV as objective function. However, if
relationships between simple attributes and NPV can be found, they are helpful to reduce CPU time in real well placement
optimizations. Using attributes have already been suggested by Artus et al. [8], but their approach still needs several
hundreds of reservoir simulations, which correspond to 520% of total objective function evaluations. Here, we propose a
quick approach based on simple attribute and correlation calculation to maximize NPV by reducing the number of reservoir
simulations to only several to several tens.
First, we study the relations between the attributes presented above (PInum and PIfield) and NPV. Figure 11 presents the
cross-plots between these attributes and NPV by random selection of 160 wells. The attribute PIfield shows a better
correlation (R2 = 0.81) with NPV than the attribute PInum (R2 = 0.47), especially for high NPV values. Although introducing
other attributes or using combination of attributes by principle component analysis can define a better relationship between
attributes and NPV, here, we use only the PIfield, which has been used in above studies, to show the efficiency of this quick
approach.
In the following, we plan to implement a well of maximum length 1000 m in the reservoir described above by
maximizing NPV in 5 years. An active aquifer is presented to support pressure. The NPV is calculated by taking into account
oil and gas prices with discounting rate, as well as drilling and operation costs. To get the reference solution, 5 optimization
runs were performed with genetic algorithm using NPV as objective function. Each optimization run takes about 5 days. The
maximum NPV value among all these runs is 99.6 MM$, and this value is considered as the reference solution.
In the section "numerical comparisons" presented above, we optimised well position with PIfield as objective function.
Each optimization run takes about 1 2 hours. Now, we plot NPV in Figure 12 for all best wells optimised with the genetic
algorithm and CMAES with population size 9 and 60. Generally, wells with high PIfield values have also high NPV. Well
configurations with high PIfield (points inside the black circle in Figure 12) are shown in Figure 13. The reference well
configuration is also shown in this figure. It is found that all well configurations are not only localised in a small region close
to the reference case, but also oriented in almost the same direction. All these wells provide qualitatively the indication about
the best well position.
In general, we can only afford 1 or 2 optimization runs in real case. To quantify the best well position in terms of NPV, it
is not sufficient using only the best well in terms of PIfield in a particular optimization run. However, we can choose several
or several tens well configurations with high PIfield and then evaluate the NPV for all these wells. This technique can
improve the well placement in terms of NPV without increasing much CPU time. In the cross-plot shown by Figure 12, one
well configuration with high PIfield value fails to provide high NPV value (the point in the red circle in Figure 12).
However, if we choose 5 well configurations with high PIfield values in this optimization run, the NPV can be greatly
improved from 6.16 10 7 to 9.13 10 7 . The cross-plot between PIfield and NPV for these 5 wells is shown in Figure 14,
and the best well position in this optimization run is shown in Figure 13 (green line). This approach can provide NPV much
closer to the reference solution, and the run takes less than 2 hours instead of 5 days for an optimization with NPV as
objective function.
In this example, primary production is considered with the presence of an active aquifer, and the field productivity index
(PIfield), issued from fast well test simulation, can be used as a pertinent attribute for well placement to maximize NPV. But

SPE 113525

in general cases, especially in the presence of water or gas injectors, other attributes are needed to be found to reduce CPU
time in the optimization procedure.

Conclusions
In well placement optimization, a large number of parameters are involved and the objective function is usually strongly nonlinear. Evolutionary algorithms seem to be the most suitable for this kind of problems. In this paper, we have presented the
application of CMAES (Covariance Matrix Adaptation Evolution Strategy) to the well placement optimization. Like
genetic algorithm, the CMAES is based on evolutionary approach for global optimization.
The CMAES and the genetic algorithm are compared through an example. In most cases, results in both methods are
comparable. But CMAES can provide more accurate and better solution. However, the population size in CMAES has an
impact on the optimization results for well placement. The efficiencies of both genetic algorithm and CMAES depend also on
model parameters, such as discretization steps in the genetic algorithm, the step size or the learning rate in CMAES. The
impact of certain model parameters on optimization results are studied. Determining best model parameters for well
placement is important issue to improve the accuracy and efficiency of evolutionary algorithms. Further studies are needed
on the sensitivity issues.
Direct optimization of well positions to maximize NPV is usually prohibitive due to large number of reservoir simulations
required in NPV calculation. An alternate quick approach based on correlation between simple attributes and NPV is
proposed to reduce CPU time in optimization procedure. In the primary production period, the field productivity index
(PIfield), which can be estimated from one step well test simulation, is a good attribute to help well optimization in terms of
NPV.

Acknowledgement
The author wishes to thank A. Wamy for programming, and M. Schoenauer and A. Auger for helpful discussions.

References

1. Fogel, L.J., Owens, A.J. and Walsh, M.J. : "Artificial Intelligence through Simulated Evolution", John Wiley, 1996.
2. Holland, J.H. : "Adaptation in Natural and Artificial Systems", University of Michigan Press, Ann Arbor, 1975.
3. Rechenberg, I : "Evolutionsstrategie - Optimierung technischer Systeme nach Prinzipien der biologischen Evolution", FrommanHolzboog Verlag, Stuttgart, 1973.
4. Koza, J.R. : "Genetic Programming: On the Programming of Computers by Means of Natural Selection", MIT Press, 1992.
5. Bittencourt, A.C. and Horne, R. : "Reservoir development and design optimization", paper SPE 38895 presented in SPE Annual
Technical Conference and Exhibition, San Antonio, Texas, 5-8 October 1997.
6. Gyagler, B., Horne, R. Rogers, L. and Rosenzweig, J.J. : "Optimization of well placement in a Gulf of Mexico waterflooding", SPE
Reservoir Evaluation and Engineering, June 2002, pp229-236.
7. Yeten, B., Durlofsky, L. and Aziz, K. : "Optimization of nonconventional well type, location and trajectory", SPE Journal, September
2003, pp200-210.
8. Arthus, V., Durlofsky, L., Onwumalu, J. and Aziz, K. : "Optimization of non-conventional wells under uncertainty using statistical
proxies", Computational Geosciences, n4, dec. 2006, pp389-404.
9. Hansen, N. and Ostermeier, A. : "Adapting Arbitrary Normal Mutation Distributions in Evolution Strategies : The Covariance Matrix
Adaptation", conference on Evolutionary Computations (ICEC '96), 1996, pp312-317.
10. Hansen, N. and Ostermeier, A. : "Completely derandomized self-adaptation in evolution strategies", Evolutionary Computation, 9(2),
2001, pp159-195.
11. Hansen, N. : "The CMA Evolution Strategy : A Tutorial", Nov. 2005.
12. Michalewicz, Z. : "Genetic Algorithms + Data Structures = Evolution Pregrams", Springer Verlag, New York, 1996.
13. Eiben, A.E. and Smith, J.E. : "Introduction to Evolutionary Computing", Springer, 2003.
14. Auger, A. : "Contributions thoriques et numriques l'optimization continue par algorithmes volutionnaires", thse doctorale de
l'universit Paris 6, 2004.

Appendix
In case of small population size, the covariance matrix can be estimated using the information in previous generations as
shown in the following formula:

C ( g +1) =

1
1
C(i +1)
g + 1 i = 0 (i )

SPE 113525

To assign recent generations a higher weight, exponential smoothing is used. Choosing C (0) = I to be the unity matrix
and a learning rate 0 ccov 1 , then C(g+1) is given by:

C ( g +1)

= (1 ccov )C ( g ) + ccov
= (1 ccov )C

(g)

(g)2

+ ccov

C g +1

w
i =1 i

xi(:g+1) m ( g )

xi(:g+1) m ( g )

(g)

(g)

The above formula is called rank--update. In particular case for = 1, we get rank-one-update. The rank-one-update is
used to construct the evolution path using exponential smoothing:
pc( g +1) = (1 cc ) pc( g ) + cc (2 cc ) eff

x1(:g+1) m( g )

(g)

where pc R n is the evolution path, 0 cc 1 is learning rate. The above summation is referred to as cumulation.
Coimbining the rank--update and the cumulation, we get an adapted covariance matrix:
T
c
C ( g +1) = (1 ccov )C ( g ) + cov pc( g +1) pc( g +1) +

cov

ccov (1

cov

rank one update

wi (

i =1

xi(:g +1) m( g )

(g)

)(

xi(:g +1) m( g )

(g)

rank update

Another notion introduced in CMAES is the step size control, which is also calculated recursively. Detailed formulae are
given in Hansen [11].

Table 1Numerical PI (PInum) simulated with genetic algorithm


Genetic algorithm
max OF
N evaluation
49660

600

53988

650

50338

3650

60982

4825

60982

2275

42873

175

60982

1650

58176

2100

56976

3375

60982

1050
Average

55594

2035

SPE 113525

Table 2Numerical PI (PInum) simulated with CMAES


=9
max OF

= 20
max OF

54779

N evaluation
2730

37379
48979

= 40
max OF

60785

N evaluation
4813

64458

N evaluation
4996

222

59841

4788

60940

4946

169

60807

3831

64036

4972

59122

2990

23347

2756

60758

4900

34638

25

60110

3627

64391

4999

53266

99

15912

3548

52540

733

56854

2608

59852

3732

60803

4993

58327

3057

64037

4993

60385

4880

18225

1927

61003

3728

64463

4983

18808

2133

59847

4975

21001

4097

1596

52554

4079

57378

max OF
64263

N evaluation
4857

max OF

60758

N evaluation
4890

64200

N evaluation
4866

61044

4903

59582

4971

58460

4871

60456

4970

60694

4950

63707

4970

60471

5000

63987

4896

58655

4984

57902

5000

60902

4988

60368

4786

60303

4809

58204

4944

60412

4731

59182

4996

60893

4760

64160

4967

54813

4819

57264

4565

63979

4623

61084

4916

63906

4996

50178

4716

64372

4940

59066

5000

60454

Average
44038

Average

= 60
max OF

4450

= 80

Average
60039

Average

= 100

Average
4924

60876

4565
Average

4893

60457

Table 3Field PI (PIfield) obtained with genetic algorithm

4808

10

SPE 113525

Genetic algorithm
max OF
N evaluation
169,49

1125

169,49

675

172,41

450

158,73

850

161,29

200

166,67

675

161,29

425

169,49

1025

166,67

275

169,49

650
Average

166,50

635

Table 4Field PI (PIfield) obtained with CMAES


=9
max OF
175,44

= 20

N evaluation
235

max OF
158,73

= 40

N evaluation
347

max OF
169,49

N evaluation
885

166,67

42

161,29

488

169,49

1069

166,67

207

161,29

693

169,49

1536

158,73

460

158,73

569

166,67

1343

175,44

465

147,06

365

169,49

1536

163,93

308

175,44

904

175,44

1601

166,67

241

166,67

528

169,49

1879

163,93

223

169,49

1201

169,49

1069

172,41

394

166,66

1037

175,44

1601

163,93

389

166,67

690

169,49

1879

296

163,20

682

170,40

Average
167,38

Average

Average
1440

SPE 113525

11

= 60
max OF

= 80

= 100

N evaluation
1911

max OF
166,67

N evaluation
934

max OF

169,49

166,67

N evaluation
2146

169,49

2065

161,29

2181

166,67

1722

169,49

2008

169,49

2017

166,67

1532

169,49

1689

175,44

2379

169,49

2007

166,67

1615

166,67

2713

166,67

1114

169,49

2364

166,67

1796

169,49

3444

169,49

2207

166,67

811

175,44

2525

175,44

2090

169,49

2003

166,67

1826

169,49

1397

169,49

1263

169,49

1835

175,44

1997

166,67

1026

175,74

2274

1934

167,85

1712

169,30

Average
170,40

Average

Average
2043

12

SPE 113525

105

104

103

102

Layer 3

101

Layer 2

Layer 1

Layer 4
Figure 1Permeability field

objective function

Figure 2Objective function in CMAES

Layer 5

SPE 113525

13

Figure 3Best well positions obtained with genetic algorithm (PInum)

=9
Figure 4Best well positions obtained with CMAES (PInum)

= 60

14

SPE 113525

Figure 5Best well positions obtained with genetic algorithm (PIfield)

= 60
Figure 6Best well positions obtained with CMAES (PIfield)

SPE 113525

15

70000

60000

PInum

50000

40000

30000

20000

10000

0
0

214

223

232

214

250

26-1

27-2

28-3

29-4

-6
210-5 211

Discrttisation
Genetic
Algorithm GA2n

212-7

213-9 214-11

Figure 7Impact of discretization in the genetic algorithm

n=4

n = -3
Figure 8Well positions in sensibility study with the genetic algorithm GA2n

n = -13

215-13

16

16

SPE 113525

70000

60000

50000

PInum

40000

30000

20000

10000

= 0.3 & multiply Ccov by 2

multiply Ccov by 2

default model parameters


0
0

91

2
20

3
40

4
60

95

6
20

7
40

8
60

99

10
20

11
40

12
60

population size

Figure 9Impact of model parameters in CMAES

default parameters ( = 20)

learning rate Ccov multiplied by 2 ( = 20)

Figure 10Well positions in sensitivity study with CMAES

13

17

9,00E+07

9,00E+07

8,00E+07

8,00E+07

7,00E+07

7,00E+07

6,00E+07

6,00E+07

R = 0,8105
NPV

5,00E+07
4,00E+07

4,00E+07
3,00E+07

2,00E+07

2,00E+07

1,00E+07

1,00E+07

0,00E+00

R2 = 0,4697

5,00E+07

3,00E+07

0,00E+00

20

40

60

80

100

120

140

160

180

5000

IP
champs
PIfield

10000

15000

20000

IP num
PInum

Figure 11cross plot NPV and attributes

1,20E+08

1,00E+08

8,00E+07
CMAES (l=9)
NPV

NPV

SPE 113525

CMAES (l=60)

6,00E+07

GA
Reference value

4,00E+07

2,00E+07

0,00E+00
155

160

165

170

175

PIfield

Figure 12cross plot for "best wells" obtained by maximising PIfield

180

25000

18

SPE 113525

Reference well position


CMAES ( = 9)
CMAES ( = 60)
Genetic Algorithm

Case surrounded with the red circle


in Figure 12 CMAES (
= 9):
Before amelioration
After amelioration

Figure 13Well configurations with high PIfield values

1,20E+08

1,00E+08

NPV

8,00E+07

6,00E+07
Wells in an optimisation run

4,00E+07

Reference value
2,00E+07

0,00E+00
162

165

168

171

174

PIfield

Figure 14Cross plot of selected wells in an optimisation run

177

Das könnte Ihnen auch gefallen