Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Automatic History Matching in a Bayesian

Framework, Example Applications


Fengjun Zhang, SPE, ChevronTexaco Energy Technology Co.; Jan Arild Skjervheim, U. of Bergen;
A.C. Reynolds, SPE, U. of Tulsa; and D.S. Oliver, SPE, U. of Oklahoma

Summary model constructed from static data is available and is represented


The Bayesian framework allows one to integrate production and by a multivariate Gaussian pdf. Then, the a posteriori pdf condi-
static data into an a posteriori probability density function (pdf) tional to production data is such that calculation of the maximum
for reservoir variables (model parameters). The problem of gen- a posteriori estimate or generation of a realization by the random-
erating realizations of the reservoir variables for the assessment of ized-maximum-likelihood method is equivalent to the minimiza-
uncertainty in reservoir description or predicted reservoir perfor- tion of an appropriate objective function.2,4–6
mance then becomes a problem of sampling this a posteriori pdf to History-matching problems of interest to us involve a few thou-
obtain a suite of realizations. Generation of a realization by the sand to tens of thousands of reservoir variables and a few hundred
randomized-maximum-likelihood method requires the minimiza- to a few thousand production data. Thus, an optimization algo-
tion of an objective function that includes production-data misfit rithm suitable for large-scale problems is needed. Our belief is that
terms and a model misfit term that arises from a prior model nongradient-based algorithms such as simulated annealing and the
constructed from static data. Minimization of this objective func- genetic algorithm are not competitive with gradient-based algo-
tion with an optimization algorithm is equivalent to the automatic rithms in terms of computational efficiency. Classical gradient-
history matching of production data, with a prior model con- based algorithms such as the Gauss-Newton and Levenberg-
structed from static data providing regularization. Because of the Marquardt typically converge fairly quickly and have been applied
computational cost of computing sensitivity coefficients and the successfully to automatic history matching for both single-phase-
need to solve matrix problems involving the covariance matrix for and multiphase-flow problems.7–10 No multiphase-flow example
the prior model, this approach has not been applied to problems in considered in these papers involved more than 1,500 reservoir
which the number of data and the number of reservoir-model pa- variables. For single-phase-flow problems, He et al.8 and Reynolds
rameters are both large and the forward problem is solved by a et al.2 have generated realizations of models involving up to
conventional finite-difference simulator. 12,500 reservoir variables by automatic history matching of pres-
In this work, we illustrate that computational efficiency prob- sure data. However, they used a procedure based on their gener-
lems can be overcome by using a scaled limited-memory Broyden- alization of the method of Carter et al.11 to calculate sensitivity
Fletcher-Goldfarb-Shanno (LBFGS) algorithm to minimize the ob- coefficients; this method assumes that the partial-differential equa-
jective function and by using approximate computational stencils tion solved by reservoir simulation is linear and does not apply for
to approximate the multiplication of a vector by the prior covari- multiphase-flow problems.
ance matrix or its inverse. Implementation of the LBFGS method For multiphase-flow problems, the Gauss-Newton method is
requires only the gradient of the objective function, which can be not efficient in its classical form because it requires the generation
obtained from a single solution of the adjoint problem; individual of sensitivity coefficients. Wu et al.9 and Li et al.10 used the
sensitivity coefficients are not needed. We apply the overall pro- adjoint method12,13 to calculate sensitivity coefficients. However,
cess to two examples. The first is a true field example in which a if the number of production data to be history matched exceeds
realization of log permeabilities at 26,019 gridblocks is generated 100, calculation of all sensitivity coefficients by the adjoint
by the automatic history matching of pressure data, and the second method is not computationally feasible. On the other hand, the
is a pseudofield example that provides a very rough approximation gradient-simulator method introduced into the petroleum engineer-
to a North Sea reservoir in which a realization of log permeabilities ing literature by Anterion et al.14 is not computationally feasible if
at 9,750 gridblocks is computed by the automatic history matching the number of reservoir variables to be estimated or simulated is
of gas/oil ratio (GOR) and pressure data. greater than 100. Although reparameterization methods such as
zonation,15,16 grad zones,17,18 pilot points,19,20 spectral decompo-
Introduction sition, and subspace methods21–24 are sometimes used to reduce
The Bayes theorem provides a general framework for updating a the number of sensitivities calculated and the number of variables
pdf as new data or information on the model becomes available. estimated directly, no model reparameterization is applied for the
The Bayesian setting offers a distinct advantage. If one can gen- problems considered in this work.
erate a suite of realizations that represent a correct sampling of the For large-scale history-matching problems in which Nd (the
a posteriori pdf, then the suite of samples provides an assessment number of production data to be history matched) and Nm (the
of the uncertainty in reservoir variables. Moreover, by predicting number of reservoir variables to be simulated or estimated) are
future reservoir performance under proposed operating conditions both large, the only viable gradient-based optimization algorithms
for each realization, one can characterize the uncertainty in future are those that do not require the direct calculation of all sensitivity
performance predictions by constructing statistics for the set of coefficients. Three algorithms that have this characteristic are well
outcomes.1,2 Liu and Oliver3 have recently presented a comparison known: truncated Newton, variable-metric or quasi-Newton, and
of methods for sampling the a posteriori pdf. Their results indicate nonlinear conjugate gradient (CG).
that the randomized-maximum-likelihood method is adequate for In the truncated Newton method, the explicit computation of
evaluating uncertainty with a relatively limited number of samples. the Nd × Nm sensitivity-coefficient matrix, G, is not required. At
In this work, we consider the case in which a prior geostatistical each iteration of the method, a matrix problem is solved by the CG
method. This solution requires only the product of G times a vector
and the product of the transpose of G times a vector; the entries of
G itself are never computed. A procedure for computing these
Copyright © 2005 Society of Petroleum Engineers
matrix vector products without first computing G was introduced
This paper (SPE 84461) was first presented at the 2003 SPE Annual Technical Conference into the petroleum engineering literature by Chu et al.,25 although
and Exhibition, Denver, 5–8 October, and revised for publication. Original manuscript re-
ceived for review 31 March 2004. Revised manuscript received 18 January 2005. Paper
the basic idea appeared much earlier in the geophysics literature by
peer approved 27 January 2005. Mackie and Madden.26 Although computation of the matrix prod-

214 June 2005 SPE Reservoir Evaluation & Engineering


ucts can be made relatively efficient using a gradient simulator run 1
to calculate G times a vector and a single adjoint solution to O共m兲 = 共m − muc兲T CM−1共m − muc兲
2
calculate GT times a vector, the CG method may require up to Nd 1
iterations to obtain convergence if the matrix is poorly conditioned + 共d − duc兲T CD−1共d − duc兲, . . . . . . . . . . . . . . . . . . . . . . . (1)
and no good preconditioning matrix is available.27 Because it 2
seems unlikely to us that a good preconditioning matrix can be where m denotes an Nm-dimensional column vector of reservoir
constructed without direct knowledge of G, we have not applied variables (in our examples, entries of m represent gridblock values
this algorithm. of log permeabilities), muc represents an unconditional realization
The CG and quasi-Newton methods are attractive optimization generated from the prior model, and duc is obtained by adding
algorithms for large-scale history-matching problems in that they noise based on production-data-measurement errors to the ob-
do not require the calculation of the sensitivity of each production served historical production data, dobs. CM denotes the Nm×Nm
datum to the reservoir variable; instead, they require only the prior covariance matrix, and CD denotes the Nd×Nd covariance
gradient of the objective function at each iteration. The gradient matrix for production-data-measurement errors. The N d -
of the objective function can be calculated by solving a single dimensional column vector dobs contains as entries all production
adjoint problem. data to be history matched. Minimization of Eq. 1 gives a single
Makhlouf et al.28 applied the CG method to estimate 450 grid- history-matched reservoir model. To generate multiple plausible
block permeabilities by history matching production data obtained reservoir models for the purpose of evaluating uncertainty, one
under multiphase-flow conditions. In one example, 110 iterations generates a sequence of (muc, duc) and, for each pair, minimizes the
were required to obtain convergence. In the second example con- objective function of Eq. 1. In each example considered here, only
sidered, 222 iterations were required to obtain convergence. It one history-matched model is generated.
appears that the authors did not use preconditioning. If a suitable Zhang and Reynolds29 have shown that the LBFGS algorithm
preconditioning algorithm could be found, it is likely that the is an appropriate optimization algorithm for large-scale history-
convergence rate could be accelerated. As discussed in Zhang and matching problems. With the algorithm, the search direction is
Reynolds,29 we have not found a preconditioning matrix or a se- given by
quence of preconditioning matrices that yield a CG optimization
algorithm that is as robust and computationally efficient as the pk+1 = −H̃k−1gk, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (2)
quasi-Newton method implemented by Zhang and Reynolds29 and where H̃k−1 is the approximation to the inverse of the Hessian
used for the history-matching examples presented later in this paper. matrix and gk is the gradient of the objective function O(m). Here,
Quasi-Newton methods have been used previously in automatic k is the iteration index. The approximation to the inverse of the
history-matching problems by Yang and Watson,30 Masumoto,31 Hessian is updated from iteration to iteration using information on
and Savioli et al.32 While these studies generally found that a changes in the gradient and the model from previous iterations.
self-scaling variable-metric method is more robust and computa- Computer memory can be saved by using information from only
tionally efficient than the CG, steepest-descent, and standard un- some of the preceding iterations to update the approximate inverse
scaled BFGS algorithms, the number of reservoir variables esti- Hessian, which leads to the term “limited memory.” The main
mated was less than 25 in all examples considered. advantage of this method as developed by Nocedal35 is that we
The more recent computational experiments of Deschamps need to explicitly compute and store only the initial approximation
et al.33 suggest that the most-efficient optimization method will be H̃0−1. At subsequent iterations, H̃k−1 is never explicitly computed
a hybrid scheme, and they specifically advocate schemes that com- or stored; instead, the search direction is computed directly. More-
bine a Gauss-Newton method with another procedure. They reject over, aside from the multiplication of the initial gradient by H̃0−1,
the Levenberg-Marquardt scheme as relying too heavily on the no matrix multiplications are required to update the approximation
steepest-descent method and do not present any comparisons based to the inverse Hessian; only vector inner products are needed in the
on this method. On the other hand, some results (see, for example, computations. It is well known36 that scaling can improve the
Li et al.10) indicate that a modified Levenberg-Marquardt algo- convergence rate of quasi-Newton methods. In our examples, we
rithm can be superior to the Gauss-Newton method when initial use scaling at all iterations on the basis of the scaling factors
data mismatches are very large; if changes in the model are not recommended in Zhang and Reynolds.29
damped at early iterations of the Gauss-Newton method, Li et al. In our application, the initial Hessian inverse approxi mation
found that excessive roughness may be introduced into the esti- H̃0−1 is chosen as the prior Nm×Nm covariance CM. Each iteration
mated (or simulated) rock-property fields. of the LBFGS algorithm requires multiplication by CM. Each it-
For the two history-matching examples they considered, Des- eration also requires evaluation of the objective function to check
champs et al.33 stated that a pure quasi-Newton method was not for convergence. To evaluate the objective function of Eq. 1 re-
competitive with the hybrid schemes they advocated. However, it quires calculation of
is not clear which quasi-Newton method they used, how they
initialized the inverse Hessian approximation, or whether they x = CM−1共m − muc兲. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (3)
used scaling. Moreover, a hybrid method that uses a classical
For problems in which the number of nonzero entries of CM is not
Gauss-Newton method at some point in the optimization algorithm
too large, the vector x is computed by solving
would not be computationally efficient for large-scale history-
matching problems because it would require the calculation of too CMx = 共m − muc兲 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (4)
many sensitivity coefficients.
by a preconditioned CG method with the level-zero incomplete
In the examples presented here, we use the scaled LBFGS
LLT decomposition of CM used as a preconditioner.37 However, for
algorithm to perform automatic history matching. As shown by
many problems of interest to us, the number of nonzero elements
Zhang and Reynolds,29 LBFGS appears to be a promising algo-
of CM are so large that the storage of all nonzero entries of CM
rithm for large-scale history-matching problems.
and/or the solution of Eq. 4 is not feasible. In such cases, we
approximate the multiplication by CM and CM−1 by using compu-
Automatic History Matching With the LBFGS tational stencils.38–40 In this approach, the equation CMCM−1 ⳱ I
Optimization Algorithm is approximated by the discrete convolution of two stencils; that is,
The typical objective function minimized in the history-matching
process includes a production misfit term and a model misfit term 兺 *兺 sten sten
−1
= I1, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (5)
that provides regulation. With the randomized-maximum- where I1 is a stencil with unity at its center point and other entries
likelihood method,6,34 a realization of the model conditional to equal to zero. The elements of ∑sten are generated by the under-
observed production data, dobs, is generated by minimizing the lying covariance function from which CM is generated. The size of
following objective function: the stencils is kept small by assuming that the stencils are station-

June 2005 SPE Reservoir Evaluation & Engineering 215


ary in an infinite domain and, thus, ignoring the effect of reservoir values obtained from well tests. A total of 104 buildup pressures
boundaries. Moreover, the approach assumes uniform grids. Even from 40 wells were used in the history-matching procedure;
with these assumptions, it appears that the method is efficient and buildup pressures used to condition the prior geostatistical model
robust only for exponential covariance functions. For the expo- were obtained at different times during the period from April 1991
nential case, very small computational stencils are required. For to January 1998. In the history-matching procedure, we specified
the Oseberg pseudofield example considered in this work, a 7×7×5 the monthly oil rate for each well on the basis of historical data.
inverse stencil was used. The covariance matrix for the prior model for horizontal log per-
meability was generated from a spherical variogram with ranges in
Tengiz Field Example the x, y, and z directions, respectively, equal to 6,560, 4,920, and
165 ft. The variance of horizontal log permeability was equal to
We consider history-matching buildup data from the Tengiz res-
0.225. After doing a history match based on this geostatistical
ervoir, a carbonate reservoir located in the Pri-Caspian basin. As
model, we tried a second history match based on the use of a
shown in Fig. 1, the central portion is relatively flat, with localized
modified geostatistical model in which we increased the correla-
structural highs on the southern and eastern edges; see Chambers
tion length (variogram range) to 730 ft in the z direction. With the
et al.41 and He and Chambers42 for additional details on the geology.
original variogram model, the vertical dimension of some simula-
Tengiz is an undersaturated-oil reservoir produced by 44 wells
tion gridblocks is almost as large as the vertical range of 165 ft. In
with initial pressure equal to 11,950 psi at a datum depth of 14,765
this situation, we expect that the horizontal log-permeability model
ft subsea. With very rare exceptions, all flowing bottomhole pres-
will exhibit little correlation in the z direction. Because no indi-
sures have been maintained above bubblepoint pressure, which is
cation of the accuracy of the pressure data was provided, we used
equal to 3,586 psi. Our history match of static-pressure data from
the observed pressure data to define duc in Eq. 1. To define CD,
pressure-buildup surveys is based on an upscaled reservoir model
of Tengiz. The upscaled reservoir model of Tengiz was created by pressure-measurement errors were modeled as independent, iden-
removing most of the sloping flanks near the outer edges of the tically distributed normal random variables with zero mean and
reservoir and upscaling the remainder of the reservoir to a variance equal to 1 psi2.
59×49×9 grid. In the upscaled model, the gridblock sizes in the x Figs. 2 and 3 show the field oil rate (dot-dashed curve) at the
and y directions are almost uniform, with values between 815 and bottom of each figure, the observed buildup pressures (stars), the
825 ft. Gridblock sizes in the z direction are nonuniform, with bottomhole pressures predicted from the initial reservoir model
values varying between 15 and 150 ft. The initial permeability and (solid curve), the bottomhole pressure predicted from the condi-
porosity fields are from a geostatistical model. tional realization obtained from history matching with short cor-
relation length in the z direction (dashed curve), and the bottom-
History Matching. He and Chambers42 have previously history hole pressure predicted from the conditional realization with long
correlation length in the z direction (dotted curve) for wells T-15
matched pressure data from the Tengiz reservoir. Their history-
and T-102, respectively. Note that the conditional realizations ob-
matching procedure, however, is quite different from the one used
tained by history matching predict pressures in much closer agree-
in this work. In their results, they conditioned the permeability
ment with the observed pressures than with those pressures pre-
field in the neighborhood of a well to well-test-pressure data by
dicted with the initial model. Fifteen iterations of the LBFGS
estimating a single permeability multiplier. The multipliers for
algorithm were required to obtain convergence.
each well were then interpolated to the whole field. The final
In most cases, the initial model of Tengiz results in predicted
model was obtained by applying the multiplier field to the initial
wellbore pressures much lower than the measured-pressure data,
permeability field generated from the geostatistical model. In our
and in several instances, these predicted pressures are several hun-
work, we match all measured pressure data simultaneously and
dreds of psi below the corresponding measured-pressure data. In
adjust log permeability in each grid cell individually during the
such situations, we expect that permeability will need to be in-
history-matching process without using permeability multipliers.
creased significantly in the neighborhood of a well to obtain a
In the history-matching problem, we used the scaled LBFGS
history match. In rare cases, the forward run of the initial model
algorithm as implemented by Zhang and Reynolds29 to modify the
predicts wellbore pressures above the observed bottomhole pres-
horizontal log-permeability field. Porosity values were held fixed
sures; see, for example, Fig. 2. In these cases, one expects that one
on the basis of the initial geological model in the history-matching
will need to decrease permeability near the well to obtain a history
process. Vertical permeability was not estimated directly but was
match of pressure data.
required to be equal to 10% of horizontal permeability at each
Both of the reservoir models obtained from the conditional
gridblock. Thus, adjustments to horizontal permeability made to
realizations with short and long correlation length in the z direction
match pressure data resulted in adjustments to vertical permeabil-
predict pressures (dashed and dotted curves, respectively, in Figs.
ity. Well skin factors were fixed at each well on the basis of the
2 and 3) that are in reasonable agreement with the measured

Fig. 2—Bottomhole pressure and oil rate of Well T-15, located in


Fig. 1—Structural map of the Tengiz field. gridblock (43,27).

216 June 2005 SPE Reservoir Evaluation & Engineering


Fig. 3—Bottomhole pressure and oil rate of Well T-102, located Fig. 4—The initial-guess (left) conditional realization (right) of
in gridblock (39,7). horizontal log permeability conditioned to pwf, based on a short
correlation length in the z direction, Layer 1.

buildup pressures, although in a few cases, pressure mismatches has been increased significantly in and near the Layer 3 gridblock
remain above 300 psi. However, many of these predicted pressures that is penetrated by well T-102 (Fig. 5). Moreover, as expected,
are much higher than those predicted with the initial rock-property the thickness-averaged horizontal permeability for Layers 1
field; for some wells, the history-matched model predicts a rela- through 3 at the areal location of Well T-102 has been increased
tively small pressure drop throughout the producing history (see significantly. This increase is consistent with the increased pres-
Fig. 3). This indicates that the history-matching process has re- sure predicted by the history-matched model (Fig. 3).
sulted in significantly higher permeabilities in regions around
these wells. On the other hand, Fig. 2 illustrates that at a few wells, Comments. Although we greatly improved the pressure match by
the history-matched model predicts pressures that are lower than history matching, we obtained a far-from-perfect match, and to
those predicted from the initial reservoir model, which indicates obtain the match, we made extremely large changes to the log-
that the history-matching process has decreased the permeability in permeability field during the history-matching process. The maxi-
some regions. mum gridblock permeability was approximately 10 md in the ini-
Figs. 4 and 5 show the horizontal log permeability from the tial model, but after history matching using the geostatistical
initial model (left plot in each figure) and the conditional realiza- model with a short correlation length in the vertical direction, we
tion of horizontal log permeability based on a short correlation obtained some gridblock permeabilities on the order of 1,000 md.
length in the z direction (right plot in each figure) for Layer 1 and The long correlation length was introduced as an experiment to see
Layer 3, respectively. The platform part of Tengiz has a higher if it would result in significant damping of the changes in the
horizontal log permeability than the flanks, especially in the top horizontal log-permeability field. As expected, the long and short
layers. Note that in much of the reservoir, the history-matching correlation-length cases resulted in very different conditional re-
process has resulted in a large increase in horizontal log perme- alizations of the log-permeability field, even though pressure
ability; although it may be difficult to tell from the figures, the matches of similar quality were obtained. However, even in the
largest changes are in gridblocks close to the well locations. For long vertical correlation-length case, we obtained some gridblock-
example, Well T-102 is located in an areal gridblock (39, 7) and permeability values on the order of 100 md, which is inconsistent
completed in model layers 1, 2, and 3. From the results of Fig. 3, with the prior geostatistical model. Moreover, there is no basis for
we see that the history-matched model predicts higher wellbore increasing the correlation length in the vertical direction.
pressures than the initial model, which suggests that permeabilities It appears that the unreasonably large changes in the log-
in gridblocks penetrated by the well have been increased by history permeability field occur because of inconsistencies between the
matching. A careful examination of the results indicates that the pressure and rate data at the wells. Rate data are based on a
permeability has actually been decreased slightly in the Layer 1 monthly average. In many cases, the “measured” buildup pressures
(Fig. 4) and Layer 2 gridblocks penetrated by Well T-102, but it correspond to periods when the well is still flowing according to
the input monthly rate data, so our automatic history-matching
procedure treats them as true flowing wellbore pressures. Because
the well is flowing at times corresponding to these measured shut-
in pressures, the optimization algorithm makes large increases in
permeability in an attempt to enable the well to flow with little or
no pressure drop. Despite these difficulties, the LBFGS algorithm
was able to generate a reasonable history match.

Oseberg Example
The second history-matching example considered is a very rough
approximation to the Oseberg reservoir, which is located in the
Norwegian sector of the North Sea. The reservoir consists of three
distinct geological zones. Etive is the top zone, and Oseberg is the
bottom zone; these two zones are separated by Rannoch, which is
a relatively low-permeability layer. There is vertical communica-
tion between the three zones. In our simulation study, we simu-
lated only one half of the reservoir using a 39×25×10 grid. In the
Fig. 5—The initial-guess (left) conditional realization (right) of simulation model, Layer 1 corresponds to Etive, Layer 2 corre-
horizontal log permeability conditioned to pwf, based on a short sponds to Rannoch, and model layers 3 through 10 are used to
correlation length in the z direction, Layer 3. model the Oseberg zone. Initial reservoir pressure is 4,071 psi at

June 2005 SPE Reservoir Evaluation & Engineering 217


the depth of 8,192 ft subsea, and the initial bubblepoint pressure is correct covariances were used. As shown in Table 1, mean values
3,771 psi at this same datum. The reservoir has a gas cap at the top of 7.48 for ln(k) and 6.47 for ln(kz) were used to generate the
and an aquifer at the bottom. The initial gas/oil contact is at 8,192 unconditional realization of all Oseberg layers. Therefore, the un-
ft subsea, and the water/oil contact is at 8,918 ft subsea. The oil conditional realization for ln(k) and ln(kz) does not have any trend
column is separated from the aquifer by a tar mat that prevents vertically in Oseberg. The means used for generating the uncon-
water coning. Fig. 6a shows the surface plot of the reservoir top, ditional realization for ln(k) and ln(kz) for the Etive layer are 6.02
and Fig. 6b shows the overview of the reservoir top, with well and 4.61, respectively. Thus, we expect that horizontal mean per-
locations indicated by white squares for producers and black meability will be decreased near the bottom of the reservoir and
squares for gas-injection wells. Note that the reservoir has a sig- increased near the top of the reservoir when history matching
nificant dip. Wells are fully penetrating. In the oil column, initial synthetic production data generated from the true model. In the
oil saturation is 0.885, and the initial water saturation is equal to history-matching process, the porosity field was fixed with the
the irreducible water saturation, 0.115. In the gas cap, the initial mean porosity in Etive, Rannoch, and Oseberg, respectively, given
gas saturation is 0.885, and water saturation is equal to 0.115. The by 0.14, 0.1, and 0.22. For the example problem, all wells are
true synthetic model for the rock-property fields was generated completely penetrating vertical wells, and production data are es-
with a sequential Gaussian cosimulation algorithm. The means and sentially insensitive to vertical permeability. Thus, vertical perme-
variances for the horizontal log permeability [ln(k)] and the verti- ability is essentially unchanged by history matching production data.
cal log permeability [ln(kz)] used to generate the truth case are The tar zone is located in gridblocks centered at (xi, yj, zk), i ⳱
summarized in Table 1. The means decrease linearly with depth in 33, 34, 1 ⱕ j ⱕ 25, and 1 ⱕ k ⱕ 10. In the tar zone, we set the
Oseberg, so we have simply recorded the means and variances for horizontal and vertical permeability equal to 1 md in the true case.
the top and bottom Oseberg model layers. Means for the interme- When history matching to generate a realization by the random-
diate model layers are obtained by linear interpolation. An aniso- ized-maximum-likelihood method, we first generate an uncondi-
tropic exponential covariance function was used for each of the tional realization muc from the prior model and then modify muc by
log-permeability fields with ranges in the x, y, and z directions, setting the entries corresponding to ln(k) and ln(kz) for the tar zone
respectively, equal to 1,968 ft, 6,555 ft, and 20 ft. In the simulation to zero.
model, we used a uniform grid (656 ft in width) in the y direction. The observed data used for history matching are constructed by
In the z direction, the eight model layers located in the Oseberg running the simulator with the true model for a total time of 2,400
zone have uniform height; ⌬z ⳱ 11.5 ft, whereas ⌬z ⳱ 23.0 ft for days. The oil-rate history for each of the five producers is shown
the Etive layer and ⌬z ⳱ 16.5 ft for the Rannoch layer. In the x in Figs. 7a through 7e. The corresponding gas-injection-rate his-
direction, the grid is uniform near the central part of the reservoir, tory for the two gas-injection wells is shown in Fig. 8. The syn-
with ⌬x ⳱ 328 ft, but as we approach the boundaries in the x thetic production data were generated on the basis of the rates
direction, ⌬x gradually increases to approximately 2,000 ft for a shown in Figs. 7 and 8.
gridblock near the boundaries. The variable gridblocks pose a In our first history-matching experiment, we history matched
problem when using stencils to approximate the action of multi- wellbore pressure from the producers and injectors and GOR data
plication by CM and CM–1 because the procedure for generating a from the producers. There were 71 measurements for each type of
stencil assumes that the grid is uniform in each direction. In this data at each well, so the total number of data history matched was
application, the stencils were based on ⌬x ⳱ 328 ft, ⌬y ⳱ 656 ft, 852. Recall that the model parameters are the horizontal log per-
and ⌬z ⳱ 11.5 ft. This approximation is expected to introduce too meability ln(k) and vertical log permeability ln(kz) in each indi-
strong a correlation between permeabilities in gridblocks which vidual gridblock; therefore, the total number of model parameters
have larger dimensions, but at this point in time, no alternative simulated is 19,500. The scaled LBFGS method was applied to
procedure is available. minimize the objective function of Eq. 1, except that no noise was
In the history-matching procedure, we used different means added to the synthetic production data. In this example, informa-
than were used in the true model when generating the uncondi- tion from the previous 30 iterations was used to construct the
tional realization, muc, of ln(k) and ln(kz); muc was used as the approximation to the inverse of the Hessian matrix at each itera-
initial guess when minimizing the objective function of Eq. 1. The tion,29 and scaling was applied at each iteration.

Fig. 6—Top depth of the reservoir and the well locations.

218 June 2005 SPE Reservoir Evaluation & Engineering


For the case in which we history match both pressure and breakthrough and higher GORs are obtained in the history-
producing GOR data, the behavior of the objective function is matched and true models than in the initial model (Fig. 11).
shown by the triangular data points in Fig. 9. Note that 46 itera- We also tried matching only pressure data and only GOR data.
tions were required to achieve “convergence,” and the objective When only pressure data were matched, the optimization algorithm
function was reduced by a factor of approximately 500, from converged to a local minimum, with many pressure mismatches on
1.4×107 to 3.2×104. The time required to generate the history the order of 50 psi; the behavior of the optimization algorithm
match is roughly equivalent to 100 reservoir-simulation runs. Figs. when history matching only pressure data is shown by the square
10 and 11, respectively, show the history match of wellbore pres- data points in Fig. 9. When only GOR data were matched, the
sure and GOR obtained at the first producing well. The top curve objective function converged more rapidly, but the same quality of
in Fig. 10 represents the wellbore pressure predicted by reservoir the GOR match was almost indistinguishable from the results of
simulation on the basis of the initial unconditional realization of Fig. 11 obtained by matching both pressure and GOR data sets.
the log-permeability fields; the bottom curve in Fig. 11 represents Interestingly, when we matched only GOR data, the model ob-
the GOR predicted by reservoir simulation on the basis of an initial tained by history matching gave predicted wellbore pressures in
model. Note that data predicted from the history-matched model reasonable agreement with wellbore pressure data; this fact is il-
(data points indicated with a plus) are in close agreement with lustrated by the results of Fig. 14. In this figure, plus data points
observed data (circular data points) but are very different from data represent pressure predicted from conditional realization of the
predicted with the initial reservoir model. Similar results were log-permeability fields obtained by matching only GOR data, and
obtained at the other producing wells. circular data points correspond to observed pressure data.
Fig. 12 shows three cross-sectional plots of the change in hori- As an experiment, we ran the reservoir simulator based on the
zontal log permeability. The results show that, as expected from true reservoir model, the initial reservoir model, and the models
Table 1, the largest increases (positive values) in permeability obtained by history matching three different data sets (pressure and
occurred near the top part of the Oseberg zone and in Etive, and the GOR, pressure only, and GOR only) to predict future performance.
largest decreases (negative values) occurred in the bottom layers of In making a future performance prediction, all well rates were
Oseberg, with the largest changes occurring in regions fairly near specified from 2,400 to 2,800 days, but after that time, each of the
the line of producing wells. Fig. 13 shows the gas-saturation pro- five producing wells was produced at a fixed constant bottomhole
files for two cross sections at 2,400 days generated with the true pressure equal to 2,000 psi. The total cumulative oil production
model, the initial model, and the permeability fields generated by obtained for each of the five models from 2,400 days onward is
history matching pressure and GOR data. Note that the gas ad- shown in Fig. 15. In this figure, the solid line, dashed line, short
vances much more slowly near the bottom of the reservoir for the dashed line, dot-dashed line, and dotted line, respectively, repre-
true and history-matched models than for the initial model. Near sent the field cumulative oil production from the true model, the
the top of the reservoir, gas advances more rapidly in the true and model obtained by history matching both pressure and GOR data,
history-matched model than in the initial model. This is consistent the model obtained by history matching only GOR data, the model
with the results of Fig. 12, as well as the fact that earlier gas obtained by history matching only pressure, and the initial model.

Fig. 7—Production-rate history of the five producing wells.

June 2005 SPE Reservoir Evaluation & Engineering 219


Fig. 9—Behavior of the objective function.
Fig. 8—Injection rate for the gas-injection wells.
gk ⳱ gradient of the objective function O(m)
The model obtained by history matching only GOR data and the G ⳱ sensitivity-coefficient matrix
model obtained by history matching both pressure and GOR data H ⳱ Hessian matrix
resulted in almost identical predictions. Note that at 3,900 days, the H̃ ⳱ approximation to the Hessian matrix
cumulative oil production obtained for the initial model overpre- I ⳱ identity matrix
dicts the cumulative production predicted from the true model by k ⳱ horizontal permeability
approximately 14 million STB, whereas the overprediction for kz ⳱ vertical permeability
each of the three history-matched models is approximately 6 mil- m ⳱ vector of model parameters
lion STB. N ⳱ dimension of vectors
O(m) ⳱ objective function
Conclusions p ⳱ search direction
The scaled LBFGS optimization algorithm was applied to perform x,y,z ⳱ coordinate directions
automatic history matching for two realistic problems in a Bayes-
ian framework. For large-scale optimization problems, the basic Subscripts
advantage of the algorithm is that individual sensitivity coeffi-
cients are not needed; each iteration of the algorithm requires only d ⳱ data (for vectors)
the gradient of the objective function, and this gradient can be D ⳱ data (for matrices)
computed from one adjoint solution. The memory and computa- k ⳱ iteration index
tional effort required is reduced further by an implementation that, m ⳱ model (for vectors)
at each iteration, uses only a limited amount of information from M ⳱ model (for matrices)
previous iterations and avoids explicit computation and storage of obs ⳱observed
the approximation to the inverse Hessian. By using discrete sten- uc ⳱ unconditional
cils to approximate the multiplication of a vector by the prior
covariance matrix and its inverse, memory requirements are re-
Superscripts
duced significantly. With the implementation of LBFGS method
used here,29 automatic history matching within a Bayesian frame- T ⳱ transpose
work is feasible for field problems. –1 ⳱ inverse

Nomenclature Acknowledgments
C ⳱ covariance matrix This work was supported by the member companies of the U. of
d ⳱ vector of data (units depend on data type) Tulsa Petroleum Reservoir Exploitation Projects (TUPREP) and

Fig. 10—Wellbore-pressure match at producing well 1. Fig. 11—GOR match at producing well 1.

220 June 2005 SPE Reservoir Evaluation & Engineering


Fig. 12—Change in log permeability in three cross sections.

the U.S. Dept. of Energy (DOE) under Award No. DE-FC26- 4. Tarantola, A.: Inverse Problem Theory: Methods for Data Fitting and
00BC15309. However, any opinions, findings, conclusions, or rec- Model Parameter Estimation, Elsevier, Amsterdam (1987).
ommendations herein are those of the authors and do not neces- 5. Oliver, D.S.: “Incorporation of Transient Pressure Data into Reservoir
sarily reflect the views of the DOE. We thank ChevronTexaco, Characterization,” In Situ (1994) 18, No. 3, 243.
ExxonMobil, Lukarco, and the Republic of Kazakhstan for allow- 6. Oliver, D.S., He, N., and Reynolds, A.C.: “Conditioning Permeability
ing us to use the Tengiz data. Fields to Pressure Data,” Proc., 5th European Conference for the Math-
ematics of Oil Recovery, Leoben, Austria (1996).
References 7. Tan, T.B.: “A Computationally Efficient Gauss-Newton Method for
1. Hegstad, B.K. and Omre, H.: “Uncertainty Assessment in History Automatic History Matching,” paper SPE 29100 presented at the 1995
Matching and Forecasting,” Geostatistics Wollogong 96, E.Y. Baafi SPE Symposium on Reservoir Simulation, San Antonio, Texas, 12–15
and N.A. Schofield (eds.), Kluwer Academic Publishers, Dordrecht, February.
The Netherlands (1997). 8. He, N., Reynolds, A.C., and Oliver, D.S.: “Three-Dimensional Reser-
2. Reynolds, A.C., He, N., and Oliver, D.S.: “Reducing Uncertainty in voir Description From Multiwell Pressure Data and Prior Information,”
Geostatistical Description with Well Testing Pressure Data,” Reservoir SPEJ (September 1997) 312.
Characterization: Recent Advances, R.A. Schatzinger and J.F. Jordan 9. Wu, Z., Reynolds, A.C., and Oliver, D.S.: “Conditioning Geostatistical
(eds.), American Assn. of Petroleum Geologists, Tulsa (1999) 149– Models to Two-Phase Production Data,” SPEJ (June 1999) 142.
162. 10. Li, R., Reynolds, A.C., and Oliver, D.S.: “History Matching of Three-
3. Liu, N. and Oliver, D.S.: “Evaluation of Monte Carlo Methods for Phase Flow Production Data,” paper SPE 66351 presented at the 2001
Assessing Uncertainty,” SPEJ (June 2003) 188. SPE Reservoir Simulation Symposium, Houston, 11–14 February.

Fig. 13—Gas-saturation profiles in two cross sections at 2,400 days predicted from the unconditional realization (left), the true
model (center), and the model obtained by history matching pressure and GOR data (right).

June 2005 SPE Reservoir Evaluation & Engineering 221


Fig. 14—Pressure at producing well 1, only GOR matched.
Fig. 15—Total field cumulative oil production, 2,400 to 3,900
days.
11. Carter, R.D. et al.: “Performance Matching With Constraints,” SPEJ
(April 1974) 187; Trans., AIME, 257.
12. Chavent, G.M., Dupuy, M., and Lemonnier, P.: “History Matching by 27. Axelsson, O.: Iterative Solution Methods, Cambridge U. Press, New
Use of Optimal Theory,” SPEJ (February 1975) 74; Trans., AIME, York City (1994).
259. 28. Makhlouf, E.M. et al.: “A General History Matching Algorithm for
13. Chen, W.H. et al.: “A New Algorithm for Automatic History Match- Three-Phase, Three-Dimensional Petroleum Reservoirs,” SPE Ad-
ing,” SPEJ (December 1974) 593; Trans., AIME, 257. vanced Technology Series (July 1993) 83.
14. Anterion, F., Eymard, R., and Karcher, B.: “Use of Parameter Gradi- 29. Zhang, F. and Reynolds, A.C.: “Optimization Algorithms for Auto-
ents for Reservoir History Matching,” paper SPE 18433 presented at matic History Matching of Production Data,” Proc., 8th European Con-
the 1989 SPE Symposium on Reservoir Simulation, Houston, 6–8 Feb- ference on the Mathematics of Oil Recovery, Freiburg, Germany
ruary. (2002).
15. Jacquard, P. and Jain, C.: “Permeability Distribution From Field Pres- 30. Yang, P.H. and Watson, A.T.: “Automatic History Matching With
sure Data,” SPEJ (December 1965) 281; Trans., AIME, 234. Variable-Metric Methods,” SPERE (August 1988) 995.
16. Jahns, H.O.: “A Rapid Method for Obtaining a Two-Dimensional Res- 31. Masumoto, K.: “Pressure Derivative Matching Method for Two Phase
ervoir Description From Well Pressure Response Data,” SPEJ (Decem- Fluid Flow in Heterogeneous Reservoir,” paper SPE 59462 presented at
ber 1966) 315; Trans., AIME, 237. the 2000 SPE Asia Pacific Conference on Integrated Modelling for
17. Bissell, R.C., Sharma, Y., and Killough, J.E.: “History Matching Using Asset Management, Yokohama, Japan, 25–26 April.
the Method of Gradients: Two Case Studies,” paper SPE 28590 pre- 32. Savioli, G.B., Grattoni, C.A., and Bidner, M.S.: “On the Inverse Prob-
sented at the 1994 SPE Annual Technical Conference and Exhibition, lem Application to Reservoir Characterization,” paper SPE 25522
New Orleans, 25–28 September. available from SPE, Richardson, Texas (1992).
18. Bissell, R.: “Calculating Optimal Parameters for History Matching,” 33. Deschamps, T. et al.: “The results of testing six different gradient
Proc., 4th European Conference on the Mathematics of Oil Recovery, optimisers on two history matching problems,” Proc., 6th European
Roros, Norway (1994). Conference on the Mathematics of Oil Recovery, Peebles, Scotland
19. de Marsily, G. et al.: “Interpretation of Interference Tests in a Well (1998).
Field Using Geostatistical Techniques to Fit the Permeability Distribu- 34. Kitanidis, P.K.: “Quasi-linear geostatistical theory for inversing,” Wa-
tion in a Reservoir Model,” Geostatistics for Natural Resources Char- ter Resour. Res. (October 1995) 31, No. 10, 2411.
acterization, Part 2, G. Verly, M. David, A.G. Journel, and A. Mare-
35. Nocedal, J.: “Updating Quasi-Newton Matrices with Limited Storage,”
chal (eds.), D. Reidell, Dordrecht, The Netherlands (1984) 831–849.
Math. Comp. (July 1980) 35, No. 151, 773.
20. RamaRao, B.S. et al.: “Pilot Point Methodology for Automated Cali-
36. Nocedal, J. and Wright, S.J.: Numerical Optimization, Springer, New
bration of an Ensemble of Conditionally Simulated Transmissivity
York City (1999).
Fields, 1. Theory and Computational Experiments,” Water Resour. Res.
(March 1995) 31, No. 3, 475. 37. Meijerink, J.A. and van der Vorst, H.A.: “An iterative solution method
for linear systems of which the coefficient matrix is a symmetric M-
21. Kennett, B.L.N. and Williamson, P.R.: “Subspace methods for large-
matrix,” Mathematics of Computation (January 1977) 31, No. 137, 148.
scale nonlinear inversion,” Mathematical Geophysics, D. Reidell, Dor-
drecht, The Netherlands (1988) 139–154. 38. Oliver, D.S.: “Calculation of the Inverse of the Covariance,” Math-
22. Oldenburg, D.W., McGillivray, P.R., and Ellis, R.G.: “Generalized ematical Geology (October 1998) 30, No. 7, 911.
subspace methods for large-scale inverse problems,” Geophys. J. Int. 39. Skjerheim, J.A. and Oliver, D.S.: “Approximation of the Inverse Co-
(July 1993) 114, No. 1, 12. variance for History Matching,” TUPREP Research Report 19, Tulsa
23. Reynolds, A.C. et al.: “Reparameterization Techniques for Generating (9 May 2002).
Reservoir Descriptions Conditioned to Variograms and Well-Test Pres- 40. Aannonsen, S.I. et al.: “Integration of 4D Data in the History Match
sure Data,” SPEJ (December 1996) 413. Loop by Investigating Scale Dependent Correlations in the Acoustic
24. Abacioglu, Y., Oliver, D.S., and Reynolds, A.C.: “Efficient Reservoir Impedance Cube,” Proc., 8th European Conference on the Mathematics
History Matching Using Subspace Vectors,” Computational Geo- of Oil Recovery, Freiburg, Germany (2002).
sciences (2001) 5, No. 2, 151. 41. Chambers, K.T. et al.: “Characterization of a Carbonate Reservoir
25. Chu, L., Komara, M., and Schatzinger, R.A.: “Efficient Technique for With Pressure-Transient Tests and Production Logs: Tengiz Field, Ka-
Inversion of Reservoir Properties Using Iteration Method,” SPEJ zakhstan,” SPEREE (August 2001) 250.
(March 2000) 71. 42. He, N. and Chambers, K.T.: “Calibrate Flow Simulation Models with
26. Mackie, R.L. and Madden, T.R.: “Three-dimensional magnetotelluric Well-Test Data To Improve History Matching,” paper SPE 56681 pre-
inversion using conjugate gradients,” Geophys. J. Int. (October 1993) sented at the 1999 SPE Annual Technical Conference and Exhibition,
15, No. 1, 215. Houston, 3–6 October.

222 June 2005 SPE Reservoir Evaluation & Engineering


try. He holds a Master of Engineering degree in industrial math-
SI Metric Conversion Factors ematics from the Norwegian U. of Science and Technology in
bbl × 1.589 873 E – 01 ⳱ m3 Trondheim. Albert C. Reynolds holds the positions of McMan
Professor of Petroleum Engineering, Professor of Mathematics,
ft × 3.048* E – 01 ⳱ m and Director of TUPREP at the U. of Tulsa, where he has been a
ft3 × 2.831 685 E – 02 ⳱ m3 faculty member since 1970. He holds a BA degree from the U.
psi × 6.894 757 E + 00 ⳱ kPa of New Hampshire, an MS degree from Case Inst. of Technol-
ogy, and a PhD degree from Case Western Reserve U., all in
*Conversion factor is exact. mathematics. He has served SPE as a Technical Editor and as
a member of the Pressure Transient Testing Committee, the
Formation Evaluation Award Committee, and the Reservoir
Monitoring Committee. Reynolds received the SPE Distin-
Fengjun Zhang is a petroleum engineer at ChevronTexaco En- guished Achievement Award for Petroleum Engineering Fac-
ergy Technology Co. e-mail: feng@chevrontexaco.com. He ulty in 1983. He became a Distinguished Member of SPE in 1999
has worked in the areas of formation evaluation, seismic data and received the 2003 SPE Reservoir Description and Dynamics
inversion, reservoir engineering, and production optimization. Award. Dean S. Oliver is the Eberly Family Chair Professor and
His interests include reservoir characterization, history match- Director of the Mewbourne School of Petroleum and Geologi-
ing, dynamic and static data integration, optimization, and cal Engineering at the U. of Oklahoma. e-mail: dsoliver@ou.
artificial intelligence. Zhang holds a BS degree in well logging edu. Previously, he was a professor in the Dept. of Petroleum
and an MS degree in applied geophysics, both from the U. of Engineering at the U. of Tulsa, as well as a research geophysi-
Petroleum China, and a PhD degree from the U. of Tulsa. Jan cist for Chevron, a staff reservoir engineer for Chevron USA and
Arild Skjervheim is currently a PhD student at the U. of Bergen’s for Saudi Aramco, and a research scientist in reservoir charac-
Centre for Integrated Petroleum Research (CIPR). He has 2 terization and optimization. Oliver holds a BS degree in physics
years of consulting experience with Scandpower Risk Manage- from the Harvey Mudd College and a PhD degree in geophys-
ment, where he performed risk analysis in the petroleum indus- ics from the U. of Washington.

June 2005 SPE Reservoir Evaluation & Engineering 223

You might also like