SPE 49004 Gradual Deformation of Continuous Geostatistical Models For History Matching
SPE 49004 Gradual Deformation of Continuous Geostatistical Models For History Matching
SPE 49004 Gradual Deformation of Continuous Geostatistical Models For History Matching
——
SPE 49004
221
2 F. ROGGERO, L. Y. HU Sm 49004
at thescale of field studies. The stochastic process introduces instance, a Iognormal distribution can be obtained with the
great discontinuity in the attempt to find a match, by following transformation:
generating a series of independent realizations. This
discontinuity finders the progressive improvement of the Y= exp(ti+ ~) .......................e..............................(2)
initial model and the efficiency of the process. where ~ and 0 are the mean and the standard deviation of
The pilot points method, proposed by de Marsily6 and
discussed in recent publications7-9, offers some solutions by lnY.
allowing continuous modifications to the geostatistical model. This formulation ensures that the geostatistical parameters
The pilot points technique consists in reducing the dimension of the new realization Z will be honored for any value of p.
of the inverse problem by constraining the initial distribution This parameterization makes it possible to continuously
to control points considered as inversion parameters. The control the generation of new models. The space of possible
pilot point values are adjusted in a deterministic optimization realizations is determined by the initial realizations ZI and
pr- to-obtain a match with dynamic data. This technique 22, when p varies between -1 and +1. For particular values of
is therefore an approximation of the complete inverse p, one of the initial realizations, or the opposite, is generated:
problem, which amounts to identiming the properties of all -Z1 is obttined for p=- 1, -22 for p=-O.5, Zl for p=O, Z2 for
the reservoir cells.”It has a certain number of disadvantages: p=O.5 and -Z1 for p=l. For other values of p, the new model
● The number and position of the pilot points are user- is a linear combination of the initial realizations with the
determined, according to criteria that are more or less same global statistical characteristics.
empirical. A change in p leads to a continuous variation in the
● The method requires constraining pilot point values by distribution Z at each grid node. Starting with an initial value
an “a priori modeI” to preserve the global statistical of p, the correlation between a new realization and the initial
characteristics of the model. distribution can be controlled by the range of variations in p.
● Finally, t~e number of pilot points can be prohibitive for For small variations, the new realization is well-correlated to
a field-scale inversion problem. the initial one. For variations equal to 0.5 or -0.5, the new
realization is completely uncorrelated to the initial one.
This paper proposes a new approach’” for gradual To increase the flexibility of the model parameterization,
deformation of the geostatistical model. The objectives of this the previous formulation can be generalized to the Iinear
approach are as folIows: combination of any number of initial realizations. Wt
. Allow reservoir engineers to modify an initial model in a {Zl,..,Zn} be a set of n independent realizations of a Gaussian
continuous way, while respecting the global statistical random field, a new realization can be expressed as follows:
characteristics.
● Propose an efficient inversion tool that can constrain a z = ~aizi ...............................................................(3)
geostatistical model to dynamic production data, by i=l
generating gi-adual improvements to the match.
where {al,..., an} are n real coefficients between -1 and 1.
Principle of the Gradual Deformation Method The geostatistical properties of the model, such as the
The basic principle of the gradual deformation method is to variogram, are preserved under the following condition:
generate a process of realizations which evolve smoothly at
n
each step. This can be achieved by introducing a correlation
ktween a new realization and the realization obtained in the x af = 1 ...................................................................(4)
i=]
pretius step. To apply this principle to continuous
geostatistical models, the solution proposed is to generate new A natural way to comply with this constraint is to use a
realizations as linear combinations of two initial models. Let variable transformation and to express the combined
21 and 22 be two independent realizations of a Gaussian realization in spherical coordinates:
random field, a new realization Z can be expressed as follows:
Z = ~COS(piZ) Z, + ~Sitl(~;Z)fiCOS(~j~) Z,+, ....(5)
Z = cos(pz)Zl i- sin(pz)Z2 .................................(1) isl j=l jzi+l
The independent realizations Z1 and 22 must be In this expression, the combined distribution Z is a
normalized, with a mean equal to zero, and unconstrained to function of n-1 independent parameters {pi,...,I..I }.
the well data. The well data can be incorporated by Realization 21 is reproduced when all parameters are equal to
constraining the combined realization Z. The mean of the mro. Otherwise, a linear combination of the n independent
actual distribution must be superimposed to the normalized realizations is obtained. The increasing number of
combined map Z, using a variable transformation. For parameters provides more flexibility in deforming an initial
222
SPE 490ti GRADUAL DEFORMATION OF CONTINUOUS GEOSTATISTICAL MODELS FOR HISTORY MATCHING 3
2
2(“ ‘:’m
‘=1
- d:”)2
.......................................(6)
1. Generation of an initial realization Zl of the error map,
Where: which corresponds to the initial guess of the-optimization
procedure.
d,ti aresimulated production results,
2. Generation of a new independent realization Z2 of the
doh aremeasured production data, error map as a search direction for an improved model.
w are weighting coefficients, 3. Initialization of p to O, to reproduce the initial guess at
nOb$is the total number of measurements. the first iteration of the optimization procedure.
4. Computation” of a new error map Z by linear
Gradually deforming the model realization results in a combination of independent realizations ZI and Z2.
corresponding gradual evolution of production results 5. Constraining the combined realization to the well data.
obtained by numerical fluid flow simulation. Consequently, 6. Transformation of the error map to take into account
the objective finction is continuous with respect to parameter the mean and the variance of the distribution. A Iognormal
variations. transformation can be applied for example.
7. Numerical fluid flow simulation and computation of the
Example. The smooth behavior of the objective finction with objective function F.
respect to the parameter variations is illustrated using the 8. Minimization of the objective function F using a
field case described in detail in the last section. The model is non-linear optimization algorithm, and computation of an
composd of 12 layers, each comprising 27x51 grid cells. improved estimate of the parameter p.
The shape uncertainty of the reservoir top structure is 9. Updating the model in an iterative procedure (i.e. return
described by a Gaussian random field. The Gaussian error to step 4) until a convergence criterion is reached.
map is superimposed over a mean map to define the actual top
depth distribution. This optimization procedure can be generalized to the
Five wells are perforated in the reservoir. A synthetic inversion of several parameters. Afier defining an initial
production history is defined from a reference realization, realization ZI, a set of new independent realizations
including prediction records of the oil rate at surface {~,...,}.} can be produced. The additional realizations
conditions and static pressures at well locations. provide more flexibility for constraining the initial model, by
A set of realizations is generated by combining two introducing several directions of perturbation. In that case,
independent initial realizations ZI and Z2. In expression (1) the parameters {pl ,...,pl.l } are initialized to zero to start from
parameter p is sampled between -1 and 1, in constant steps of the initial guess. The optimization procedure then allows
0.05, to cover an entire period. A numerical fluid flow simultaneous identification of the n-1 parameters, in order to
simulation is performed and the objective function is minimize the objective function.
Other deterministic parameters can also be matched
223
4 F. ROGGERO, L. Y. HU SPE-49004
simultaneously with the inversion of the realization itself. solution obtained in the previous step.
The field case described in the last section illustrates this A continuous chain of constrained realizations can be
possibility: the stochastic model which describes the shape of obtained by the following procedure, illustrated in Figure 2:
the top structure is constrained simultaneously with the
inversion of the activity coefficient of the aquifer. Another 1. Simulation of initial realization ZI.
example is the inversion of the mean of the distribution 2. Simulation of a new independent realization ZZ.
combined with the improvement of the error map itself. 3. Initialization of p (combination parameter) to zero.
Use of the gradient methodl *“13significantly increases the 4. Optimization of the model by combining images ZI and
efficiency of the optimization algorithms. The gradient Z2, to minimize the objective function E
method implemented in the numerical fluid flow simulator
Z = cos(pm)Zl + sin(pz)Z2
computes the derivatives of production results with respect to
model parameters. These parameters usually correspond to 5. Replacement of the initial realization Z1 by the optimal
uniform permeabilityy or porosity values over given reservoir model, Z.
zones. In ~adual deformation, the model is a linear 6. Return to step 2 until a satisfactory match of dynamic
combination of a set of initial realizations. An extension of data is obtained.
the gradient method calculates the derivatives of, production
results with respect to model deformation parameters. This This results in a Markovian prmess in which each new
method was implemented for permeabilityy or porosity constrained realization depends only on the previous step.
distributions. The objective finction decreases continuously during the
For other types of parameters, gradients are calculated by process. Consecutive optimization either allow to improve
numerical approximations. In this case, the optimization the match or preserve the match obtained in the previous step
algorithm is initialized by launching a series of numerical when p is zero. Note that a new realization can be totally
simulations corresponding to small perturbations in the independent of the previous step when the optimal p reaches
inversion parameters. The gradients are then calculated by an absolute value of 0.5.
linear regression using all the available simulations. The Renewing the search directions at each step of
procedure allows simultaneous use of two methods: the optimization leads theoretically to an overall optimum. In
gradient method for certain parameters and numerical practice, the number of optimization is limited by stopping
derivatives for others. the process when a convergence criterion is reached. This
criterion can be defined by a Iimit value for the objective
Process of ,Constrained Realizations finction, or when the rate of decrease drops below”a threshold
The gradual deformation of an initial geostatistical model value.
makes it possible to efficiently enhance the matching of Various strategies can be adopted to create a chain of
dynamic production data. The optimization method constrained realizations. A first variant is obtained by
implemented is a purely deterministic process, for a given set generalizing the previous procedure to combinations of any
of initial realizations. number of realizations at each step. A variant of a 4-image
The limited number of parameters used to control the chain is shown in Figure 3. Increasing the number of
model deformation allows to explore only a reduced subset of parameters improves the convergence rate, thereby reducing
the space of possible realizations. Of course, the number of the length of a chain, at the price of a higher cost for each
parameters can be increased to offer greater flexibility in the optimization. The bst compromise must therefore be found
choice of a new realization, However, too many parameters to obtain the minimum number of numerical simulations.
can hinder the efficiency of the optimization algorithm or lead A different strategy can be obtained by optimizing in
to prohibitive calculation times, consecutive steps. This gives the pyramid structure illustrated
The search for an optimal model by combining a limited in Figure 4. The procedure is initialized by generating a
set of realizations can lead to an unsatisfactory production series of independent unconstrained realizations. The first
data match. An improvement of the result can be attempted optimization phase combines these images two by two in
by repeating the optimization procedure using new order to obtain a new series of independent realizations, with
independent realizations, in a random process. improved matches in comparison to the initial realizations.
The major di~culty of this type of prccess is to avoid This procedure is successively reiterated until the final model,
losing the benefit of the match obtained at each step. Random corresponding to the apex of the “pyramid”, is obtained. A
re-initialization of all the initial realizations generates number of other variants can be generated by combining these
discontinued behavior of the objective function and the model different approaches and varying the number of parameters in
deformation. -10 create a ContlmIOUS
“ “ - Ot“ realizations
chain “” the optimization procedure. Nevertheless, the basic principle
constrained by successively improved matches, the proposed is maintained: to create a continuous chain of constrained
solution is to initialize each new step with the optimal realizations by successively inputting new search directions.
224
SPE 49004 GRADUAL DEFORMATION OF CONTINUOUS GEOSTATISTICAL MODELS FOR HISTORY MATCHING 5
Synthetic reservoir description. The synthetic model The measurements are weighted to obtain an equivalent
illustrated in Figure 5 is a 2D reservoir composed of a single effect regardless of the pressure level. The weighting
layer with a constant thickness of 10 m. The permeability coefficients are defined by a fraction of the value of each
distribution is Iognormal, the rn.ean permeability is 200 mD measurement, with a threshold of 0.01 bar:
and the standard deviation is 100 mD (the standard deviation
of the permeabilityy log is 0.472). The variogram is spherical o~b~= 0.01x max(~~,,O.01) ..................................(8)
with a correlation length of 480 m in hth directions. Other
petrophysical properties are constant: porosity is 0.25 and For the observation wells, measurements corresponding to
total compressibility is 5.104 bar-*. The flow is single phase periods less than 10 hours were eliminated. The objective
with a viscosity value of 1 cP. finction is normalized by the total number of observations.
A producing well with a radius of 7.85 centimeters and no This definition corresponds to a satisfactory match for all
skin is positioned in the center of the reservoir. Its history wells, when the objective function is less than 0.5.
comprises a period of constant production rate at 100 m3/day Different variants of gradual inversion chains were tested,
during 2.106 seconds, and a period of pressure buildup (zero by varying the number of independent simulations combined
rate) during the same amount of time. Four observation wells at each step of optimization. The results obtained by
are located around the producing well, regularly spaced at combining 2, 5 and 10 maps are compared in the follow-up
640 m. study. The gradient method was used to improve the
The geostatistical simulation grid is regular, composed of performance of the optimization algorithm.
51 cells with a 40 m side in each direction. For numerical
fluid flow simulations, the gridding was progressively refined l-parameter chain. The process implemented is illustrated
locally until regular cells of 0.4 m were obtained around the in figure 2. A random realization is used to initialize the
producing well. chain. The value of the objective function for the first
Synthetic dynamic data were defined by a numerical flow numerical simulation is 878. A second independent
simulation run based on the reference model. The production realization is simulated. A first optimization step with one
data comprise the bottomhoIe pressure evolution at the parameter produces an optimal combination of two
producing well and four observation wells (see Figure 5). independent realizations. The objective function decreases to
219 within four iterations.
Inverse problem. me permeability distribution of the To carry on the minimization of the objective function, the
reference model, shown on Figure 5, is assumed unknown. initial realization is replaced by the optimal medel and a new
The objective is to build new realizations of the model while independent search direction is generated. A new 1-parameter
preserving global statistical properties and by integrating optimization step once again improves the objective function.
dynamic data. The process continues until a satisfactory match is obtained or
The inversion process is initialized by simulating a stops when the objective function decreases at a rate that is
random realization. Statistical parameters such as the mean too low. The convergence criterion is defined by a threshold
and variogram are assumed to be known. New realizations value of 0.5 for the objective function.
are not constrained to the wells. To constrain the model and Figure 6 shows the evolution of the objective finction and
reduce uncertainties we implemented a gradual deformation the optimal combining parameter p during the process. A
chain to simultaneously match the pressures to the five wells. good convergence rate is observed at the beginning of the
The match was obtained by minimizing the objective function chain: the objective function decreases from 838 to 16 in eight
in the following way: optimization. The performance strongly deteriorates after
that: the objective function continues to decrease but the
1 ‘~’ P:m – <:, 2 convergence rate is very low. The process is stopped afier
F=—
x(
....................................(7)
2nob, i=~ 100 optimization, with a final value of 1.35 for the objective
‘:bs ) function. Convergence was therefore not reached, despite
strong improvement in the matching.
The evoIution of optimal parameter p during optimization
(Fig. 6) reveals significant variations at the beginning of the
6 F. ROGGERO, L. Y. HU SPE 49004
chain, then lower or zero values at the end of the process. zones. Averaging of 50 constrained realizations; shown on
These variations correspond to improvements in the model. Figure 11, highlights a strongly permeable zone in the
When p reaches an absoIute value of 0.5, the model is de- vicinity of well 03 and a zone with lower permeability
correlated fiorrt the preceding phase and the objective between the producing well and well 04.
function is strongly improved. On the contrary, no The model is therefore highly constrained by the
improvement is obtained for a vaIue of zero. interference test. The contribution of the dynamic data can be
To improve the convergence rate, the number of linear quantified by calculating the distribution of uncertainties
combination realizations was successively increased, first to 5 before and after matching. Standard deviation maps
and then 10. The number of independent parameters to be calculated on normalized Gaussian distributions, shown on
optimized aIso increased to 4 and 9, respective y. Hgure 1I, reveal that uncertainties clearly decrease in the
center of the reservoir, in those areas located between each
Comparison of different strategies. A same initial observation well and the producing well (i.e. in the “area of
realization was used to test the different alternatives. investigation” of the test).
Convergence rates obtained by processes using 1,4 and 9 The uncertainties on the permeability values at the well
parameters are compared in Figure 7. locations also decrease strongly. Table 1 shows that the
Increasing the number of parameters largely improves the permeability at the producing well was accurately reproduced,
rate of conm~=--—With 4 parameters, convergence is and that uncertainties on values at observation weIls were
reached in 40 optimization. With 9 parameters, only 4 strongly reduced.
optimizatiotis are rquired. These results can be explained by
@ater flexibility in model parameterization. Full Field Application Case
Excellent performances are therefore obtained with 9 The second application of the gradual deformation method is
parameters. The reader should note that at this stage no a semi-synthetic field case. The reservoir geometry and
benefits are gained from increasing the number of parameters petrophysical properties are derived from actual data but the
as any further optimization is proportionally more costly. In production history used for the inversion is simulated. The
order to compare performances, the total number of numerical history match, conducted for a IO-year period, invoIves the
fluid flow simulations per variant was plotted on Figure 7. simultaneous inversion of the top map and the aquifer
To demonstrate the effectiveness of the method we activity.
attempted to assess the number of numerical simulations
required to obtain a constrained realization in a random Reservoir description. The reservoir, shown on figure 12,
prccess. Starting with the”same initial realization, a series of covers a surface area of 2700 m by 5100 m. It is bounded to
1000 independent realizations were simulated. A highly the east and south by an active aquifer (modeled by an
discontinuous evolution of the objective function is shown on analytical aquifer) and by a major fault in its Western part.
Figure 8. For all 1000 realizations, only five models The oil-bearing zone is delimited vertically by the water oil
_wded to an objective function ranging from 30 to 40. contact and the top structure. The numerical simulation grid
Furthermore, the values did not generate a satisfactory match comprises 12 layers, each containing 27x51 cells.
of dynamic data. In this case, it was therefore impossible to The main producing zone is located in the center and
randomly find a matched modeI out of 1000 realizations. comprises three producing wells (wells LU1, LU2 and LU6).
An additional oil zone, of minor extension, is produced by a
9-parameter chain. The 9-parameter strategy was used in fourth well (well LU3), in the South West of the reservoir.
the rest of the study. The process was repeated with new Two other (inactive) peripheral wells are located in the
initial distributions to generate a set of 50 constrained water-bearing zone to the North (well LU4) and at ‘the edge of
realizations. The average cost of the method is about 40 the fault (well LU5).
numerical simulations per inversion chain. The recovery mechanism is driven by the aquifer activity.
Figure 9 shows an example of a match produced for the The reservoir has no initial gas cap and pressure is
first realization. The numerical simulation performed from maintained above the bubble point during the entire duration
the unconstrained initial realization is very far from of production.
observations. The match was obtained tier a chain of 4 The objective of this study was to quandfy the
Optim”zations. It resulted in very good match of uncertainties on the shape of the top structure and on aquifer
measurements for all 5 wells. activity. Structural uncertainties are modeled by geostatistics.
Figure 10 shows a series of 5 constrained realizations. The shape of the top structure was obtained by superposing
These realizations are different but all lead to the same degree the realization of an error map over a mean top structure, as
of dynamic data matching. A comparison with the reference in the diagram presented on Figure 13. The error map is
distribution in Figure 10 shows that the major trends have described by the following statistical parameters:
been reproduced with good positioning of the most permeable . Zero mean.
226
.- ,+.
SPE 4ti - GRADUfi “DEFORMATION OF CONTINUOUS GEOSTATISTICAL MODELS FOR HISTORY MATCHING 7
—-—r.
scattering is observe~w~e lnltlaI-realizations with an References
uncertainty of about 5570. The reader should note that the 1. Deutch, C.: “Conditioning reservoir models to weII test
lowest aquifer activity coefficient values result in the shut- information”, in Soares, A. (cd.), Geostatistics Troia92, pp.
down of the wells at the end of the simulation period. To 505-518, Kluwer Academic Pub., 1993.
isolate the effect of structural uncertainties, the forecasts
2. Hegstad, B.K., Omre, H., Tjelmeland, H. and Tyler, K.:
“Stochastic simulation and conditioning by annealing in
obtained tier matching the aquifer activity coefficient are reservoir description”, in Armstrong, M. and Dowd, P.(eds.),
also presented in Figure 19. The uncertainties decrease when Geostatistical simulations, pp. 43-55, Kluwer Academic Pub.,
the top structure maps are matched: the reference simulation 1994.
is reproduced with an error of about 590 (F]gure 20). This 3. Mitchell, M.: “An introduction to genetic algorithms;’, The MIT
can b explained by the production mechanism used and by Press, 1996.
the low uncertainties on the Initial Oil In Place in the 4. Alabert, F.: “Constraining description of randomly
drainage area of the wells. heterogeneous reservoirs to pressume test data a Monte ‘Carlo
study”’, paper SPE 19600 presented at the SW AnnuaI
Technical Conference and Exhibition, San Antonio, October
Conclusions
1989.
1. The gradual deformation method makes it possible to 5. Oliver, D.S., Cunha, L.B. and Reynolds, A.C.: “Markov chain
continuously modi~ the geostatistical model while respecting Monte Carlo methods for conditioning a permeability field to
its global characteristics such as the variogram and the mean. pressure data”, Math. Geology, 29( 1)61-91, 1997.
2. Coupled with an optimization algorithm, the gradual 6. Marsily de, G. and others: “Interpretation of interference tests
deformation..me~d~ “ve tool for constraining a in a well field using geostatistical techniques to fit the
geostatistical model with dynamic production data. A permeability distribution in a reservoir model”, in VerIy, G. et’
eonstrained realization is obtained by successive al. (cd.), Geostatistics for natural resources characterization,
improvements of an initial model in a continuous process. Part 2, 831-849, D. Reidel Pub. Company, 1984.
3. Different strategies can be used to obtain optimal 7. Wen, X.H., Deutsch, C.V., Cullick, A.S.: “High resolution
Reservoir Models Integrating MuItiple-Well Production Data”,
efficiency, by selecting the number of deformation parameters paper SPE 38728 presented at the SPE Annual Technical
in the model and the optimization sequences. Conference and Exhibition, San Antonio, October 1989.
4. In comparison with other techniques, such as simulated 8. Bissell, R.C., Dubrule, O.. Lamy, P., Swaby, P., and Lepine,
annealing or random processes, the number of numerical o.: “Combining Geostatistical Mcdelling With Gradient
sirm.dations is significantly reduced. Information for History Matching The Pilot Point Meth@”,
5. The methcd was successfiIly applied to interference test paper SPE 38730 presented at the SPE Annual Technical
interpretations and to a reservoir history match. In both Conference and Exhibition, San Antonio, October “1989.
cases, a match of all the we~ production data was obtained for 9. Roggero, F.: Tlrect Selection of Stwhastic Model Realizations
Constrained to Historical Data”, paper SPE 38731 presented at
different equiprobable realizations of the model. A decrease
the SPE Annual Technical Conference and Exhibition, San
in uncertainties by integration of the geological description
Antonio, October 1989.
and dynamic data was observed. 10. Hu, L.Y. and Helios Reservoir Group: “Gradual deformation
6. The purpose of this paper focuses on the global and iterative calibration of Gaussian-related stochastic models”,
deformation of continuous stochastic models in a Gaussian paper submitted to Mathematical Geology, 1998.
framework. However, the method was generalized to the 11. Anterion, F., Eymard, R., Karcher, B.: “Use of Parameter
deforrnatim. of any kind of Gaussian-related stochastic Gradients for Reservoir History Matching”: SPE18433,
models. The calibration of a facies model to pressure data is Houston, TX, 6-8 Feb. 1989.
described by Hu and B1anc14. 12. Rahon, D., Blanc, G., Gu&rillot, D.: “Gradients Method
Constrained by GeoIogicaI Bodies for History Matching”, SPE
36568, enver, CoIorado, 6-9 October 1996.
Acknowledgments
13. Roggero, F., Gu6rillot, D.: “Gradient Methd and Bayesian
This work was conducted in the framework of the HELIOS Formalism Application to Petrophysical Parameter
joint project between Elf Exploration Production (EEP) and Characterization”, 5’hEuropean Conference on the Mathematics
IFP. The authors wish to thank EEP and IFP for permission of Oil Recovery, Leoben, Austria, 3-6 Sept. 1996.
to publish the results and = for providing field data. The 14. Hu, L.Y., Blanc, G. and Helios Reservoir Group: “Constraining
synthetic interference test case was built in collaboration with a Reservoir Facies Model to Dynamic Data Using a Gradual
Mokhles Mezghani (IFP) who used this data to develop the Deformation Method”, 6’h European Conference- on the-
pilot points method in the framework of his thesis. Mathematics of Oil Recovery, Peebles, Scotland, 8-II Sept.
1998.
The ATHOS numerical fluid flow simulator was used for
the full field case. It wti developed jointly by IFP and
BEI~ ~. The SIMTESTW numerical well test
simtdator was initially de~eloped by ~ and is now upgraded
in the framework of the WELGEM consortium.
228
SPE 49004 GRADUAL DEFORMATION OF CONTINUOUS GEOSTATISTICAL MODELS FOR HISTORY MATCHING 9
Table 2- Field case: initiai Oil In Place and aquifer coefficient for 12
realizations, before and after History Matching
.
-.—. ..—
~Objective Function
A
L
20
10
k
I o 1
I -1 -0.75 -0.5 -0.25
P
0 0.25 0.5 0.75 1
— ,..
I “a
—
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
. . . .—.——— -—
Fig. 1- Continuous evolution of the objective function and of the model realization (normalized error maps constrained to well data) with respect
to the parameter p, va~[ng ‘between -1 and +1 by constant steps of 0.05. A numerical simulation is performed for every model realization.
229
10 F. ROGGERO, L. Y. HU s= 49004
I Z1 I I Z2
I I
Z3 1 I Z4 I
w’
I
Iwl
Z5
I 27 [
l-l
Z5
I
Zlo
1 I
*
214[ 215
-..
I I
. . —-. .—— *
——-
+rp
.—-.—-
recess of constrained realizetiona Fig. 3- Generalization of the process of constrained rea~zetions to
the optimization of severai parameters at each step
4 4
I I r I
m
.
Well 03
Producer P
~iw
o 1----
m. [.)
lm2003GQ400m
Permeability
Well 04
0.9
08
~;;
04
[ 03
.“.
-02
0.1
“ E
0 !-2www-4cQ00w o l--mm
%0[8] Thn. (s)
Well 01 Well 02
0,B
08
g~’;
0,4
\ 0,3
02
0?
0 o
w
Tlwl. (s)
. .
Fig. 5- Interference test caae: definition of the reference permeability distribution used for producing the synthetic pressure data.
230
SPE 4- GRADUAL DEFORMATION OF CONTINUOUS GEOSTATISTICAL MODELS FOR HISTORY MATCHING “11
. -----
lm
L. lo-
t
8 ,.
!
t -0,
I
.
0 10 m m40m 6oma0mtw 0 10 20 30 40 sow 70 m m fou
brebn brtibn
—-—
Fig. 6- Evolution of me objective function and of the parameter p Fig.. 7- Influence of the number of parameters on the convergence
—. durfng tie inversion process rate (objective function evolution)
104 I
o Im m m .im ml m m mo mo Iom
mti,tik.m
Fig. 8- Discontinuous behavior of the objective function obtained with a random process
an - . . . ,. an - 14
‘!
1 r + MP1!
aa aa
e.1 a7.
as ,,,
,.,’
i a5
,..’ ‘
Q4 04 ,,,”
!: 1: ,,”
w a3 ,,,
02 az ,,”
,’
a~
,’
,’ \
0
0 --- 4x0xc 0 -- — UOxu 0 - - - -
-w - *) mmw)
+
,. .... . .,.
u
-.wci
—tia
mu
,
,..’
,A.,
..
,,,”
P
.’-\,. ,..”
,,/“”- .,,
,,~ ,..’ ‘i. .,”
,.,
,,,’ ,,,
,,’
,“
,’
,’
0
0 -- - -
- *I
Fig. 9- initial simulation results and pressure match obtained at the 5 wells with the first realization.
The intarpreteUon of the buildup flow period at rna p~ucer is also presented (bottom rfght)
—-”.
F. ROGGERO, L. Y. HU 3PE 49004
12 ——
Permeability
I I
100 200 300 400 500 0.5 0.75 1.0 1.25
Fig. 11- Average permeability of the 50 constrained realization (left) and uncertainty distributions: a priori (middle) and a posterior (right).
The gray scale on the right represents the standard deviation computed with the lnK normalized distributions
232
SPE 49004 GRADUAL DEFORMATION OF CONTINUOUS GEOSTATISTICAL MODELS FOR HISTORY MATCHING 13
—
Fig. 12- Full field case: structural definition and posltlonlng of wells
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 -1900 -1800 -1700 -1600
Ffg. 13- Top depth In the referance case. A normalized error map (left) is superimposed
onto the mean depth of the top structure (middle) to obtain the final definition (right)
233
?4 F. ROGGERO, L. Y. HU SPE 49004
m+ Q
!
110 I
t
lCO 1
0 m Iaolmmmm -m
Thlle (*)
- 1.5
I - 1.0
Fig. 15-7 realkatlons of the top structure constrained to the production data (normalized error maps), compared to the reference case (top left)
234
WE 4@ GRADUAL DEFORMATION OF CONTINUOUS GEOSTATISTICAL MODELS FOR HISTORY MATCHING 15
~m
180>
0 Oala
. . -W
-------- -
.. .. —Mth
140. ...
-.. . .
.. .
120.. ...
.$ .. .
100
.... lW . ..
-.. .
. . . . ..
20 ..$ w- ....
m.
..$
.. . .... .. . .
‘t ........... I
1
40. 40
20. 20
Well LUI Well LU2
0? 01 I
o m imiwmzmm mm o m lmlw2m2mm m40co
Tlma (dqa) Time (days)
m —-—.
-—
——.
——.—...—.
---— ..—-.
. -... m
m m
..”. —.--—
-—----’--—--- -------——-----------—---.--”-------------—--
~1 ml
~.
“.’%:’”.. i
7W
t
Ico
t Well LU1
Fig. 16- History match example obtained at two wells for the realization #4
1.2
I
...%–
~,~,:
1.0
0.8
0.6
0.4
0.2
Fig. 17- Top depth uncertainty distributions before and after History Mstching (standard deviations comDuted from the
12 realizations of the normalized error m%ps)
235
76 F. ROGGERO, L. Y. HU s= 49004
.—., —-c= —..
., -
..—.
4.5-
z 4“ “
~ 3.5 --
&
= 3- -
2.5- -
$
8 2- -
i.5- -
3
1- -
.- “t 0.5- -
0
0 1000 2000 3000 4000 5000 6000
Time (days)
— ---
Fig. 18- Production forecasts obtained with the 12 initial realizations
-..
4.5-
F 4“ -
@ 3.5 --
g
= 3- -
2.5- -
$
=
.. 02
1.5- -
.- 3
e ‘- -
; 0.5 --
o-t
o 1000 2000 3000 4000 5000 6000
Time (days)
r- —.
Fig. 19- ffnce~inty reduction obtained by constraining the aquifer coe~cient to the production data
4.5- 1
F 4“ -
* 3.5 --
$
0
= 3- -
2.5- -
i
=
02
1.5. -
1
z “ -
~ 0.5- -
03
0 1000 2000 3000 4000 5000 6000
Tim e (days)
-. . .—,—-. .. . . . =.- -.
Fig. 20- Production forecasta obtained with the 12 constrained realizations
236