Вы находитесь на странице: 1из 10

A Newton-Raphson Iterative Scheme for

Integrating Multiphase Production Data


Into Reservoir Models
Zhan Wu, SPE, Texas A&M U.

Summary
This paper presents a new linearized iterative algorithm for building reservoir models conditioned to multiphase production data
and geostatistical data. The significant feature of the proposed
algorithm is that the computation of the sensitivity coefficients of
production data, with respect to model parameters, can be avoided.
This leads to dramatic reduction of computational cost. Instead of
generating the sensitivity matrix as required for the least-squares
algorithms, the proposed approach relies on solving the inversion
equations, which are derived from the necessary conditions for a
functional extremum.
It is proved that the proposed method requires considerably less
computational effort than the traditional algorithms such as the
Gauss-Newton method, the Levenberg-Marquardt method, and the
generalized pulse-spectrum technique (GPST). This is because the
computation of the sensitivity coefficients makes the traditional
algorithms computationally intensive. However, for the proposed
linearized iterative scheme, the computational requirement only
depends on the timesteps used in the reservoir simulator rather than
the number of parameters or the number of observed data.
The linearized iteration scheme converges quickly because the
inversion equations can be solved through the Newton-Raphson
method. At each iteration, the new approach requires solving the
finite difference equations and the linear adjoint equations only
once, respectively. Since the solver for the flow equations can be
used to solve both the adjoint equations and the inversion equations, the proposed algorithm can be easily applied to commercial
reservoir simulators. In this paper, two numerical examples for
incorporating water oil rate data into geostatistical models are
given to prove the efficiency of the proposed algorithm.
Introduction
Integrating production data into reservoir models is an inverse
problem and requires efficient optimization algorithms to minimize the least-squares objective function. In the past two decades,
many optimization methods were proposed for the integration. In
general, these optimization methods can be classified into three
categories: the gradient-based methods, the global optimization
methods, and the sensitivity coefficient-based methods. As early as
the 1970s, the gradient-based methods were used in automatic
history-matching problems. For single-phase flow, Chen et al.1 and
Chavent et al.2 proposed the optimal control method to calculate
the gradient of the objective function with respect to model parameters and used the gradient-based optimization methods to minimize the least-squares objective function. Later, Wasserman et al.3
extended the optimal-control method to automatic history matching of a pseudosingle-phase reservoir. Watson et al.4 estimated
water and oil relative permeabilities based on displacement experiment data. In the groundwater area, Carrera and Neuman5 used the
optimal-control method for solving the parameter identification
problem of groundwater flow. Sun and Yeh6 have also applied similar procedures to estimate aquifer properties. Yeh7 presented a
clear survey on parameter identification in the groundwater area.
Copyright 2001 Society of Petroleum Engineers
This paper (SPE 74143) was revised for publication from paper SPE 62846, first presented
at the 2000 SPE/AAPG Western Regional Meeting, Long Beach, California, 1923 June.
Original manuscript received for review 28 June 2000. Revised manuscript received
26 February 2001. Manuscript peer approved 26 March 2001.

September 2001 SPE Journal

Readers can refer to the latest review published by Ewing et al.8 on


reservoir parameter estimation, including some discussion of
optimization algorithms. In the mathematical community, Ito and
Kunisch9 and Kunisch and Tai10 used the system equations as
constraints and introduced dual variables to study the parameter
estimation problem.
For the gradient-based methods, the linear search for the optimal step size will dominate the inversion process. For the global
optimization methods, low convergence rate is the major
disadvantage. The numerous simulation runs for evaluating the
objective function become computationally expensive as the reservoir model size increases. Normally, it is preferable to use the
sensitivity coefficient-based methods such as the Gauss-Newton
method, the Levenberg-Marquardt method, and the least squares
QR decomposition (LSQR) method for minimization because of
their fast convergence. However, if a large number of production
data and reservoir parameters are involved, estimating the sensitivity
coefficients of production data with respect to the reservoir parameters is time-consuming and costly. Reducing the cost for computing the sensitivity coefficients can significantly reduce the computational cost of the integration process. In previous research work,
researchers have made great efforts to develop possible algorithms
to obtain the sensitivity coefficients. For single-phase flow,
Carter's method is the most efficient for calculating the sensitivity
coefficients.11 Unfortunately, Carter's method cannot be applied to
calculate the sensitivity coefficients for multiphase flow. The
GPST method proposed by Tang et al.12 can only be used to
calculate the sensitivity coefficients of pressure with respect to permeability.13 Recently, the optimal control method to generate the
sensitivity coefficients by Wu et al. is very efficient if the number
of observed data is less than the number of reservoir parameters.14
If reservoir models contain a few parameters, and the number of
reservoir parameters is less than the number of observed data, the
gradient-simulator method is preferable. Obviously, neither the
optimal-control method nor the gradient-simulator method for
calculating the sensitivity coefficients is efficient if the number of
observed data and reservoir parameters are large.
In this paper, unlike previous research in which the primary
focus is to develop the algorithm for estimating the sensitivity
matrix, a linearized iterative scheme for integrating production data
into geostatistical models is proposed. The feature of the proposed
algorithm is that the nonlinear inversion equations are derived from
the necessary conditions for a functional extremum rather than the
normal equations formulated from the sensitivity matrix and, thus,
the calculation for the sensitivity coefficients can be avoided. For
updating reservoir parameters once, the computational requirement
is to solve the adjoint equations n times, in which n is equal to the
number of timesteps used in the reservoir simulator. The computational cost of the proposed linearized iterative approach is only
dependent on the number of timesteps rather than the number of
observed data or the number of model parameters.
The proposed linearized iteration method converges quickly
because the inversion equations can be solved through the NewtonRaphson method. Moreover, the proposed iterative scheme offers
two other advantages. First, the linearized iterative algorithm is
easier to implement, because the sensitivity coefficients are not
required. Second, the adjoint equations are linear and have the
same coefficient matrix structure as the flow equations. No specific solver for adjoint equations is required. One can apply the same
343

solver used in the reservoir simulator to solve the adjoint equations. Actually, the solver can also be used to solve the inversion
equations. However, it must be pointed out that one of the difficulties in solving the inversion equations is the ill-conditioned
Jacobian matrix. Theoretically, the covariance matrix can be used
to regularize the ill-posed problem, and therefore, many algorithms
can be used to solve the inversion equations. Currently, the preconditioned conjugate gradient approach or a partial singular value
decomposition15,16 is still widely used for solving ill-posed systems. In this paper we modify Moras approach16 to solving inversion equations.
There are three major parts in this paper. First, the partial differential equations for compressible two-phase flow in reservoirs,
as well as the discrete formulation for the equations, are described
briefly, and the inverse problem for integrating production data
into reservoir models is stated. Next, the adjoint equations are
developed by means of the optimal-control theory. Then, the inversion equations are derived according to the necessary conditions
for a functional extremum. Finally, two numerical examples are
included to demonstrate some of the features of the linearized iterative schemein particular, the convergence rate for integrating
oil flow rate into geostatistical models.
Forward Model
The two-phase (water/oil) flow equations in a reservoir with spatially dependent permeability and porosity in the flow domain, W,
can be written as1719
k krw

S
(p + o g z ) = w + qw ,.

t Bw
Bw w

. . . . . . . . . . . (1)

k kro

S
(p + o g z ) = o + qo .

t Bo
o o

. . . . . . . . . . . . (2)

and
S w + So = 1 .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (3)

The reservoir can be modeled as a rectangular domain with no flow


boundary condition
p
= 0 on G .
n

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (4)

The initial conditions are given by


p ( x, y , z ,0) = p0

( x, y , z )

. . . . . . . . . . . . . . . (5)

S w ( x, y , z,0) = S wi

( x, y , z ) .

. . . . . . . . . . . . . . . (6)

The subscripts o and w=the oil and water phase. S=the saturation,
while m=viscosity. The relative permeabilities krw and kro=functions of water saturation, Bw and Bo=formation volume factors, and
qw and qo=water and oil flow rate, respectively. G =the boundary of
the reservoir. For simplicity, the capillary pressure is ignored.
Eqs. 1 and 2 are a set of nonlinear-coupled partial differential equations, which can be solved by use of finite difference methods. When
using the standard five-point scheme for the 2D reservoir, we may
write the discrete formation of the partial differential equations as
Fl n +1 = X ln +1 X ln ,

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (7)

where

Fl n+1

= f l ,1n +1 L

fl ,nM+11

fl ,nM+1

X ln+1

= xln,1+1 L xln,+M11

xln,+M1

and

denote the flow and accumulation terms, respectively. Here, n=the


timestep, and l=oil phase or water phase.
344

If the permeability and porosity are given, the solution of equations can be determined uniquely by applying the implicitpressure, explicit-saturation (IMPES) method. The performance of
the reservoir can be predicted accurately. In this study, the goal is
to obtain more realistic reservoir models by minimizing the production mismatch while the geostatistical data are honored. We
need to minimize the following objective function, which is given as
H H
H H
H
H
J o = (d d obs )T CD1 (d d obs ) + ( m m)T CM1 (m m), . . . . . (8)
to obtain the maximum posteriori (MAP) estimation of reservoir
models. CD=the diagonal covariance matrix for measured
data,

while CM=the covariance


matrix of parameter field; d contains the

calculated data. Finally, dobs=a vector of the observed data, m=the


mean of reservoir parameters, and m=a reservoir parameter vector
to be estimated. The first term in Eq. 8 represents the weighted sum
of squares of the differences between the observed data and the calculated data. The second term denotes the prior information, which
contains the mean, variance, and covariance of reservoir parameters. Oliver,20 Chu et al.,21 Reynolds et al.,22 and He et al.23 have
published a series of papers dealing with the incorporation of
production data into geostatistical data using the Bayesian estimation24 and Tarantolas inversion theory.25 More details on the
theory basis can be found in their papers.
If the Gauss-Newton method is used, the MAP solution for
minimizing the objective function Jo can be obtained by solving
the following linear equations:
H H
H
H
(G T CD1G + CM1 ) m = CD1 (d d obs ) + CM1 (m m), . . . . . . (9)
where G=the sensitivity matrix. These relationships are known as
the inversion equations or the normal equations. Note that the
inversion equations are explicit functions of the sensitivity matrix.
That is, in order to invert the parameters m, the sensitivity coefficient matrix must be evaluated in advance. As mentioned earlier,
since the governing equations are nonlinear and the number of
gridblocks is large, calculating the sensitivity matrix is not an easy
task. This forces us to develop new inversion equations to replace
Eq. 9 without introducing the sensitivity matrix G.
In the next section, the linearized iterative scheme is proposed
based on the classical optimal control theory. The process of integrating production data into geostatistical data can be considered
as a distributed parameter optimal control problem in which the
state variables (pressure and saturation) are functions of space and
time, and the control variables (permeability and porosity) are
functions of space only. However, the theory for dealing with the
distributed-parameter optimal-control problem is very tedious.
Without making approximations, it is difficult to directly study the
nonlinear distributed parameter optimal-control problem (i.e., the
nonlinear partial differential equations for multiphase flow). In
general, one can discretize the partial differential equations in both
space and time domains. The distributed parameter system is
approximated by a lumped parameter system. Then the lumped
optimal control theory can be used to analyze the distributed
parameter system. Such an approximate approach is not very rigorous in theory but very practical. For a more complete discussion
of optimal control theory, readers can refer to many books and
papers on this subject.2631 In this work, the partial differential
equations for the two-phase flow are discretized in both space and
time domains. The linearized iterative scheme is described by algebraic equations, and the matrix notation is applied throughout.
Inversion Equations
On the basis of optimal control theory, the flow equations may be
considered as constraints while the objective function J is minimized. By using the Lagrange multiplier, one can adjoin the equality
constraints (two phase flow equations) to the objective function J0 by
means of M dimensional adjoint variables, or the time-dependent
T

Lagrange multipliers ln +1 = ln,1+1 L ln,M+11 ln, M+1 . Thus,


the constraint optimal control problem becomes an unconstraint
optimal control problem. The objective function can be written as
September 2001 SPE Journal

J = J0 +

N 1

n +1 T
l

) [ Fl n +1 X ln +1 + X ln ] ,

. . . . . . . . . . . . (10)

l =o , w n = 0

where N=the total number of timesteps for solving flow equations;


the superscript T=the transpose of a matrix. The objective function
J can be expanded by the Taylor series, and the linear term is
expressed as

J J0 +

J
J
0
p n+1 + n0+1 S wn +1
n +1
l = o , w n = 0

Sw

N 1

F n +1
J
J
+ 0 k + 0 + (ln +1 )T ln +1 p n +1
k

p

F n +1
X n +1
+(ln +1 )T ln +1 S wn +1 (ln +1 )T nl +1 p n +1
p
Sw

X n
F n +1
+(ln +1 )T ll S wl + (ln +1 )T l k
k
Sw
X n +1
X n
) l + (ln +1 )T l ,

n +1 T
l

. . . . . . . . (11)

T
J T
J0
n
n
0

p
+
n Sw
n

p
S

l = o , w n = 0

N 1

J
J
+ 0 k + 0
k

F n
+( ) ln p n + (ln )T
p
n T
l

In order to satisfy Eq. 13, the variations dP and dSw must vanish.
This leads to the discretized adjoint equations
T

Fl n n X ln n X ln n +1
J0
n l n l + n l = n ,

p
p
p
p

l =o,w

. . . . . . . . . . . . . . . . . . . . . . . (14)
and
Fl n n X ln n X ln n +1
J
l n l + n l = 0n ,
n
Sw
l =o,w
w
Sw
Sw
. . . . . . . . . . . . . . . . . . . . . . . (15)

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (16)

and the adjoint equations can be solved backward in time.


Otherwise, a two-point boundary-value problem must be solved,
which is much more costly than solving the initial-value problem.
By applying Eqs. 14 and 15 to Eq. 12, the first-order variation of
the objective function becomes

J J0 +

T
J T
J0
0
+

l =o ,w n =0

N 1

F n+1
+(ln+1 )T l k (ln+1 )T
k
X n
+(ln+1 )T l .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (13)

oN 1 = wN 1 = 0,

where variations dp and dSw can be interpreted as the perturbation


of pressure and saturation due to the perturbations of permeability,
dk, and porosity, df.
After using the integration by parts, we can further rewrite Eq. 11
as follows:

J =

J = 0.

for n=1,2,,N-1. Eqs. 13 through 15 are called the first-order necessary conditions for the optimal control problem defined in Eq. 10.
It must be pointed out that Eqs. 14 and 15 are a set of ordinary
equations. The initial conditions must be required for solving the
Eqs. 14 and 15. Fortunately, we simply specify the initial condition as

X n +1
X n
) nl +1 S wn +1 + (ln +1 )T nl p n
p
Sw

n +1 T
l

The necessary condition for a minimum is that the first variation of


the objective function is equal to zero for arbitrary variations dp,
dSw (i.e., the following equation must hold):

Fl n n
n Sw
Sw

X l
X n
(ln )T ln p n (ln )T nn S wl
p
Sw
X n
X n
+(ln +1 )T ln p n + (ln +1 )T nl S wn
p
Sw

X ln+1

. . . . . . . . . . . . . . . . . . . . . . . . . (17)

We can readily obtain the gradient expressions of the objective


with respect to model parameters from Eq. 17.
T
n +1
N 1
J
J
F
= l ln+1 + 0 , . . . . . . . . . . . . . . . . (18)
k l =o , w n = 0 k
k

T
T
n
L 1
X n +1
J
J
X
= l ln +1 l ln+1 + 0 , . . . . . (19)
l =o ,w n =0

where
T

F n +1
X n +1
) l k (ln +1 )T l
k

n +1 T
l

+ (

X
+ (ln +1 )T


n
l

N
N
( Fl N AlN ) N
N T ( Fl X l )
N
+(lN )T
Sw
p + (l )
N
N

S
w

0
0
( Fl 0 X l0 ) 0
0 T ( Fl Al )
0
+(l0 )T
Sw .
p + (l )
0
0

S
w

. . . . . . . . . . . . . . . . . . . . . . . (12)

September 2001 SPE Journal

J0 J0
1 H
,

= CM ( m m ) .

. . . . . . . . . . . . . . . . . . . . . . (20)

Here, we use the fact dp0=0 and dSw0=0 because the initial values
of pressure and saturation are constant. Eqs. 18 and 19 provide the
gradient of the objective function with respect to the permeability
and porosity, and one can use the gradient-based methods to minimize the function J. However, the gradient-based methods require
estimating optimal step length, which is computationally costly.
Instead of using the gradient-based approaches, we set the gradient
equal to zero. This yields the following nonlinear equations with
respect to permeability and porosity:

F n +1 T
J 0
n +1
l
= 0,
l +
k
l =o,w n=0 k

N 1

. . . . . . . . . . . . . . . . . . (21)

345

T
X n T
X ln+1 n+1 J 0
n +1
l
= 0 . . . . . . (22)
l
l +

l = o , w n = 0

N 1

The above equations are the framework of linearized iterative


scheme. In fact, the Gauss-Newton method and the LevenbergMarquardt method are also developed based on the assumption that
the first derivative is equal to zero. The only difference is that the
proposed iterative method does not need to estimate the sensitivity
coefficients explicitly for developing the inversion equations.
For solving the nonlinear inversion Eqs. 21 and 22, one can use
the Taylor series to expand the nonlinear inversion equations

r
r
r
J J r 0
J (m0 )
2 J
r = r ( m + m) =
r + r r m +L + . . . . . . (23)
m m
m
mm
Neglecting the higher order terms and substituting Eqs. 21 and 22
in Eq. 23, the change of the reservoir parameters can be estimated
by solving the following linearized system
1

2 J J
H
m = H2
H,
m m

. . . . . . . . . . . . . . . . . . . . . . . . . . . . (24)

2 J
where H 2 =the Jacobian matrix, which can be obtained by
m
taking the partial derivatives of Eqs. 21 and 22 with respect to permeability and porosity. It is not difficult to obtain the second order
partial derivative expressions (see Appendix B):
N 1
2J

=
2
k
l = o , w n = 0 k

N 1
2J

=

2 l =o ,w n=0

F n +1 T
2 J
0
l ln +1 +
2 ,
k
k

. . . . . . . . . (25)

X n T
X n+1 T
2 J0
n +1
n +1
l
l

l
l + 2 ,


. . . . . . . . . . . . . . . . . . . . . . . (26)

N 1
J0
J
=
,
k l =o ,w n =0 k

. . . . . . . . . . . . . . . . . . . . . . . . . . (27)

N 1
2 J0
2J
=
.
k l =o , w n=0 k

. . . . . . . . . . . . . . . . . . . . . . . . . (28)

The iterative procedure for the nonlinear Eq. 24 refers to the


Newton-Raphson method. A more compact form can be obtained
by defining the Jacobian matrix as follows
2 J0
2
k
CM1 = 2
J

2J

k
,
2
J0

. . . . . . . . . . . . . . . . . . . . . . . . . . . (29)

and
A=
N 1

n +1 T n +1
2 k ( Fl ) l
l = 0, w n = 0

N 1

n T n +1
n +1 T n +1

( X l ) l ( X l ) l

l = 0, w n = 0

. . . . . . . . . . . . . . . . . . . . . . . (30)
The nonlinear inversion equations can be rewritten as
H
J
( A + Cm1 ) m = H .
m
346

. . . . . . . . . . . . . . . . . . . . . . . . . . . (31)

At each iteration, Eq. 31 can be solved to obtain the increment in


parameters dm and the model parameters can be updated at (k+1)
th iteration according to
H
H
H
m k +1 = m k + m .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (32)

k+1 can be substituted into the flow equaThe new parameters m


tions to obtain the calculated production data.
The Newton-Raphson method has high convergence rate for
well-posed problems, provided the initial values of the parameters
are close enough to the solution. If the Jacobian matrix is ill-conditioned, regularization is required. Mathematically, the inverse
matrix of the covariance matrix CM-1 is not only a weak constraint
term for the reservoir parameters but also a regularization term forcing the Jacobian matrix to be nonsingular. Therefore, many solvers
can be used to solve the inversion equations. In practice, the conjugate gradient methods are popular choices for solving the ill-posed
problem. Here, we use the preconditioned conjugate gradient
method proposed by Mora.16 In Appendix A, the main procedure of
the modified preconditioning conjugate gradient method is given.
Even though the CM-1 matrix is introduced in Eq. 31, the unique
solution of inverse problem might not be obtained. This is because
the nonlinearity of the flow equations and the least-squares objective function causes the objective function to be flat, or because
many relative minima exist in the vicinity of the minimum. In such
cases, it is impossible to obtain the minimum except when the initial guesses for the parameters are close enough to the true model.
Nevertheless, for the MAP estimation, we always use the mean of
the parameters as an initial model. In most cases, the mean is not
an accurate initial guess. As a result, a large initial data mismatch
results in an overcorrection of the models, and a physically meaningful parameter field may not be generated after just one iteration.
In order to generate smoother parameter fields when the mismatch
is large, in the previous work,14 we artificially used large variance
of observed data at the early iterations. In this work, we adopt the
Levenberg-Marquardt idea (i.e., add a term in the main diagonal of
the Jacobian matrix). The modified inversion equation is given by
H
J
( A + CM1 + I ) m = H .
m

. . . . . . . . . . . . . . . . . . . . . . . (33)

Note that the same solver used for solving the flow equations can
be applied to solve the inversion equations. Other efficient solvers
can also be used to solve Eq. 33 because the matrix A is sparse and
symmetric. In such a case, the inverse of the covariance term must
be calculated approximately because of the difficulty of computing
the inverse of the covariance matrix. In the next section, we will
show that satisfying inversion results can be obtained without
doing the inversion for the covariance matrix CM .
The main steps for the linearized iterative scheme are summarized as follows.
1. Solve the flow equation to obtain the production data at each
well using the implicit scheme or IMPES scheme, and store the
pressure and saturation data at all gridblocks on hard disk.
2. Take the partial derivatives of the flow equations with respect
to the pressure and saturation of each gridblock to formulate the
coefficient matrix of adjoint equations.
3. Solve the adjoint equations backward in time to obtain the
adjoint variables.
4. Set the gradient of the objective function, with respect to the
model parameters, to zero to obtain the nonlinear equations on
the permeability and porosity.
5. Take the first-order partial derivatives of the nonlinear inversion equations with respect to permeability and porosity to formulate the Jacobian matrix.
6. Solve the linearized inversion equations to obtain the parameter increment, dm, and update the model parameters
k+1=m
k+1+dm
.
.m
k+1 into the flow equations and solve the flow
7. Substitute m
equations. Evaluate the objective function. Check the convergence criterion. Stop if the criterion is satisfied. Otherwise,
return to Step 2.
September 2001 SPE Journal

Numerical Examples
In the following two examples, we assume that the permeability
distribution of a reservoir is log-normal, with log variance of 0.40,
and the log-mean of permeability is equal to 4.0 . The mean of
porosity is 0.20, with a prior variance of porosity equal to 0.0015.
The covariance matrix can be generated from the variogram.19 The
length of each cell is 60 ft. Fluid properties are shown in Table 1.
Fig. 1 shows the gridblocks for the 2D reservoir and the well locations. The reservoir consists of 625 gridblocks, and the number of
parameters to be estimated is 1,250 (permeability and porosity).
The total flow rate for each production well is fixed at 250 B/D,
and the injection rate at each well is fixed at 700.0 B/D. The
observed data (oil rate) are generated from the true permeability
and porosity fields by running the simulator 600 days. The total
measurement time for the observed data is between 135 and 255
days and the maximum time interval of measurement data is
5 days. Thirty-six oil rates are taken as the observed data for each
well, and the total of 334 observed data are integrated. We assume
the standard deviation of the observed data is one percentage of the
oil rate.

In the first example, the means for porosity and permeability


are selected as the initial values for the linearized iterative scheme.
Figs. 2 and 3 show the true porosity and permeability fields,
respectively, which are generated from the Cholesky decomposition of the prior covariance matrix. The MAP estimates of porosity
and permeability fields are shown in Figs. 4 and 5, respectively.
We see that the MAP estimates capture the major features of the
true cases and display their features, but the MAP estimates of both
permeability and porosity are much smoother than the true cases.
Similar results are also obtained by use of the Gauss-Newton
method.14 It is important to note that the Gauss-Newton method
requires generating the sensitivity coefficients, while the NewtonRaphson linearized iterative scheme only requires solving the
adjoint equations once for each iteration.
In the first example, the preconditioned conjugate gradient
method is used to solve the inversion equations proposed by
Mora.16 The drawback of Moras procedure is that the inverse for
covariance matrix must be estimated. In general, the size of the
parameter covariance is too large to be inverted efficiently. In
Producers
Injectors

TABLE 1RESERVOIR PARAMETERS


Parameter
Initial pressure, p
Residual oil saturation, Sor
Initial water saturation, S wc
Oil compressibility, Co
Water compressibility, C w
Water viscosity, w
Oil viscosity, o
Wellbore radius, r w
Reservoir thickness, h
Water volume formation factor, B w
Oil volume formation factor, Bo
Number of production wells
Number of injection wells

Value
4000.00 psi
0.20
0.16
-1
0.000001 psi
-1
0.0000004 psi
1.00 cp
0.82 cp
0.30 ft
20.00 ft
1.00 RB/STB
1.27 RB/STB
9
4

Fig. 1Simulation gridblocks and well locations.

13

13

19

19

25

25

13

19

25

13

19

25

0.13 0.16 0.19 0.23 0.26 0.30 2.70 3.24 3.77 4.31 4.84 5.38
Fig. 2True porosity field.
September 2001 SPE Journal

Fig. 3True permeability field.


347

order to overcome this difficulty, an approximation is made for


inverting the covariance matrix. Only the main diagonal elements
are kept, and nondiagonal elements are set to zero. Note that this
approximation is made only when the inverse of the covariance is
required. Otherwise, the true covariance matrix should be always
used. In the second example, the mean is still used as an initial
guess for the iteration. Figs. 6 and 7 show the MAP estimates of
the porosity and the log-permeability fields. Comparing Figs. 6
and 7 with Figs. 4 and 5, we can see that the MAP solutions
obtaining by use of the preconditioned conjugate method and the
modified preconditioned gradient method are quite similar. The

approximation for the inverse of the covariance matrix is valid


and will not affect the final inversion results. Mora thought that as
long as a suitable preconditioning technique is used, good inversion results could be obtained.
It was found that the uncertainty level of the permeability was
reduced significantly when the production data were honored.
However, our recent results,14 as well as those of Landa et al.,32
showed that matching the water/oil ratio can slightly reduce the
uncertainty. From the two examples, we can see that the production
data may reflect the permeability characteristic, provided sufficient
observed data are incorporated into the objective function.

13

13

19

19

25

25

13

19

25

13

19

25

0.13 0.16 0.20 0.23 0.27 0.30 2.70 3.24 3.77 4.31 4.84 5.38
Fig. 4MAP estimate of porosity conditioned to oil rate.

Fig. 5MAP estimate of permeability conditioned to oil flow rate.

13

13

19

19

25

25
1

13

19

25

13

19

25

0.13 0.16 0.20 0.23 0.27 0.30 0.13 0.16 0.20 0.23 0.27 0.30
Fig. 6MAP estimate of porosity conditioned to oil flow rate.
348

Fig. 7MAP estimate of permeability conditioned to oil flow rate.


September 2001 SPE Journal

Fig. 8 shows the objective function behavior at each iteration.


We can observe that the convergence rate of the preconditioned
conjugate gradient method is fast. In addition to the covariance
matrix, the other new preconditioning techniques should be developed to speed up the convergence rate of the proposed iterative
scheme. However, the new preconditioning technique has not been
implemented yet.
Fig. 9 shows the final match of oil flow rate at Well 1. It can be
seen that the agreement between the observed data and the calculated
data from the MAP estimates are consistently matched very well.
Conclusions
In this paper, a Newton-Raphson iterative approach for integrating
production data into geostatistical models is proposed. This algorithm does not need to calculate and store the sensitivity matrix,
because the least-squares inversion equations are derived from the
necessary conditions of a functional extremum rather than the sensitivity coefficients. This results in tremendous computational time
and storage savings over the traditional approaches. For each iteration, the new algorithm requires only one reservoir simulation run
and solving the linear adjoint equations and the nonlinear inversion
equations once, respectively. The fact that a single solver is used
to solve the flow equations as well as the adjoint equations makes
the new algorithm readily applicable to commercial simulators.
Moreover, a modified preconditioned conjugate gradient is implemented to solve the ill-posed inversion equations. It is shown that
high-resolution models can be obtained without inverting the
parameter covariance matrix. This new method can be extended to
solve large-scale integration problems.
Discussion
We are concerned about the existence and uniqueness of the solution in the Newton-Raphson iterative scheme. If the initial values
are not sufficiently close to the true model, the solution may not be
obtained, even though the covariance matrix can be used as a regularization term to remedy such a problem. In most cases, an additional regularization term is still required in additional to the
covariance matrix. However, the determination of the optimal regularization term is difficult, as no general theories exist.
We are also concerned about the convergence rate of the proposed linearized iterative scheme. It can be seen from Eq. 11 that
the objective function is expanded by the Taylor series, and only
the linear term is taken. Theoretically speaking, the new approach
is still the first-order convergence-rate method, or the so-called
pseudosecond-order convergence-rate method. In order to further
improve the convergence rate, the second-order term of the Taylor
series should be taken into account. If so, it will make the iterative
procedure more complex.
The efficiency of solving the adjoint equations plays an important role in improving the efficiency of the proposed algorithm. As
mentioned earlier, the coefficient matrix of the adjoint equations is
the same as that of the flow equations. In principle, the same solver
can be used to solve the adjoint equations. However, the coefficient

matrix of adjoint equations is nonsymmetric. The specific preconditioning techniques should be used to improve the convergence
rate if iterative methods are used to solve the adjoint equations. By
now, we use the direct and iterative methods to solve the adjoint
system, and they work very well.
Solving the adjoint equations backward in time means that the
pressure and saturation data of all timesteps at each gridblock must
be stored. This is a drawback of the proposed approach. In general, the history data can be stored on the hard disk, and no extra
computer memory will be taken. Nevertheless, reading the data
from data file may take some time. This is a tradeoff for storing the
data in the computer memory. In the near future, this disadvantage
should be overcome.
A single solver can be used to solve the flow equations, the
adjoint equations, and the inversion equations, and hence the proposed linearized iterative method can be readily extended to any
reservoir simulator. The extra computational work for the proposed
algorithm is to calculate the first-order partial derivatives of the
flow equations with respect to pressure, saturation, permeability,
and porosity, as well as the second partial derivatives of flow equations with respect to permeability and porosity. Actually, the firstorder derivatives of the flow equations with respect to pressure and
saturation will be done if the flow equations are solved by means
of an implicit scheme.
The preconditioning operator is an important factor for the
algorithm. The covariance matrix was chosen as a preconditioning
matrix because of no extra cost to the algorithm. In additional to
the covariance matrix, many techniques can be developed to generate preconditioning matrices to speed up the convergence rate of
the proposed iterative linearized method.
In this paper, two synthetic cases for integrating oil flow rate
into geostatistical models are given. The efficiency of the new
algorithm will be studied by using more cases.
Nomenclature
A = matrix
B = formation volume factor
c = compressibility
C = covariance matrix

d = vector of data
f = flow term
Fm = flux term for phase m
G = sensitivity matrix
h = reservoir thickness
I = identity matrix
J = objective function
k = permeability
= vector of model parameters
m
= mean of reservoir parameters
m
N = index of final timestep

300

100,000

Oil Flow Rate, B/D

250

Objective Function

Permeability estimation
10,000

Porosity estimation

1,000

200
150
Observed data

100

1st iteration
30th iteration

50
0
100
0

10

15

20

25

Iteration Number

Fig. 8Objective function with iteration.


September 2001 SPE Journal

30

35

50

100

150

200

250

300

Time, days
Fig. 9Oil flow-rate match at Well 1 for the first example.
349

p
Q
r
S
x
X
G
l
m
r
f
W
d

=
=
=
=
=
=
=
=
=
=
=
=
=

pressure
source/sink term
radius
saturation
accumulation term
accumulation terms
boundary of the reservoir
vector of adjoint variables
viscosity
density
porosity
flow domain
increment of reservoir parameter

Subscripts
D = data
l = phase, oil or water
M = model
obs = observed
or = residual oil
w = water phase
wc = initial water
o = oil phase
Superscripts
n = the nth timestep
T = transpose
Acknowledgments
I would like to thank Mr. T. Mathisens and Dr. K.N. Kulkarni for
their comments and their review of the manuscript.
References
1. Chen, W.H. et al.: A New Algorithm for Automatic History
Matching, SPEJ (December 1974) 593; Trans., AIME, 257.
2. Chavent, G.M., Dupuy, M., and Lemonnier, P.: History Matching by
Use of Optimal Theory, SPEJ (February 1975) 74; Trans., AIME, 259.
3. Wasserman, M.L., Emanuel, A.S., and Seinfeld, J.H.: Practical
Application of Optimal-Control Theory to History-Matching
Multiphase Simulator Models, SPEJ (August 1975) 347; Trans.,
AIME, 259.
4. Watson, A.T. et al.: History Matching in Two-Phase Petroleum
Reservoirs, SPEJ (December 1980) 521.
5. Carrera, J. and Neuman, S.P.: Estimation of Aquifer Parameters under
Transient and Steady State Conditions: 1 Maximum Likelihood
Method Incorporating Prior Information, Water Resour. Res. (1986)
22, 199.
6. Sun, N.Z. and Yeh, W.G.: Coupled Inverse Problem in Groundwater
Modeling Sensitivity Analysis and Parameter Identification, Water
Resour. Res. (1990) 26, 2507.
7. Yeh, W.G.: Review of Parameter Identification Procedures in
Groundwater Hydrology, Water Resour. Res. (1986) 22, 95.
8. Ewing, R.E. et al.: Estimating Parameters in Scientific Computation,
IEEE Computational Science Engineering (1994) 1, 19.
9. Ito, K. and Kunisch, K.: The augmented lagrangian method for parameter estimation in elliptic systems, SIM J. Control Optim. (1990) 28,
113.
10. Kunisch, K. and Tai, X.-C.: Sequential and parallel splitting methods
for bilinear control problems in Hilbert spaces, SIM J. Numer. Anal.
(1997) 34, 91.
11. Carter, R.D. et al.: Performance Matching with Constraints, SPEJ
(April 1974) 187; Trans., AIME, 257.
12. Tang, Y.N. et al.: Generalized Pulse-Spectrum Technique for 2-D and 2Phase History Matching, Applied Numerical Mathematics (1989) 5, 529.
13. Chu, L., Reynolds, A.C., and Oliver, D.S.: Computation of Sensitivity
Coefficients for Conditioning the Permeability Field to Well-test
Pressure Data, In Situ (1995) 19, 179.
14. Wu, Z., Reynolds, A.C., and Oliver, D.S.: Conditioning Geostatistical
Models to Two-Phase Production Data, SPEJ (June 1999) 142.
15. Jackson, D.D. Interpretation of Inaccurate, Insufficient, and
Inconsistent Data, Geophys. J. R. Astr. Soc. (1972) 28, 97.
350

16. Mora, P.: Elastic Wave-field Inversion of Reflection and Transmission


Data, Geophysics (1988) 53, 750.
17. Peaceman, D.W.: Fundamentals of Numerical Reservoir, Elsevier, New
York City (1977).
18. Aziz, K. and Settari, A.: Petroleum Reservoir Simulation, Elsevier,
London (1979).
19. Mattax, C.C. and Dalton, R.L.: Reservoir Simulation, Monograph
Series, SPE, Richardson, Texas (1990) 13.
20. Oliver, D.S.: Incorporation of Transient Pressure Data into Reservoir
Characterization, In Situ (1994) 18, 243.
21. Chu, L., Reynolds, A.C., and Oliver, D.S.: Reservoir Description
From Static and Well-Test Data Using Efficient Gradient Methods,
paper SPE 29999 presented at the 1995 SPE International Meeting on
Petroleum Engineering, Beijing, 1417 November.
22. Reynolds A.C. et al.: Reducing Uncertainty in Geostatistical
Description With Well-Testing Pressure Data, Proc., Fourth
International Reservoir Characterization Technical Conference,
Houston, 24 March (1997) 443462.
23. He, N., Reynolds, A.C., and Oliver, D.S.: Three-Dimensional
Reservoir Description From Multiwell Pressure Data and Prior
Information, SPEJ (September 1997) 312.
24. Gavalas, G.R., Shah, P.C., and Seinfeld, J.H.: Reservoir History
Matching by Bayesian Estimation, SPEJ (December 1976) 337;
Trans., AIME, 261.
25. Tarantola, A.: Inverse Problem Theory-Methods for Data Fitting and
Model Parameter Estimation, Elsevier, New York City (1987).
26. Sage, A.P.: Optimum Systems Control, Prentice-Hall, New York City
(1968).
27. Lions, J.L.: Optimal Control of Systems Governed by Partial
Differential Equations, Springer, New York City (1971).
28. Bryson, A.E. and Ho, Y.: Applied Optimal Control, Hemisphere
Publishing Corp., Washington, D.C. (1975).
29. Ray, W.H. and Lainiotis, D.G.: Distribution Parameter Systems, Marcel
Dekker, New York City (1978).
30. Butkovskiy, A.G.: Distribution Parameter Control System Survey,
Automation and Remote Control (1979) 40, 1568.
31. Butkovskiy, A.G.: Theory of Optimal Control by System with
Distributed Parameters, American Elsevier, New York City (1981).
32. Landa, J.L. et al.: Reservoir Characterization Constrained to Well Test
Data: A Field Example, paper SPE 36511 presented at the 1996 SPE
Annual Technical Conference and Exhibition, Denver, Colorado,
69 October.

Appendix A
This appendix presents a brief procedure of the modified preconditioned conjugate-gradient method for the nonlinear inversion
equations. For a more complete account, readers can refer to Ref. 16.
Here, we present the main steps.
For k=1, maximum iteration number,
H H
d = d cal . . . . . . . . . . . . . . . . . . . . . . . . calculated data (A-1)
H H
H
H
d k = d d obs , mk = mk m
H H
1 H H
J 0 = (d d obs )T CD1 (d d obs )
2
~
H
1 H
+ (m m)T CM1 (m m) . . . . . . objective function (A-2)
2

J J
,
gk =

pk = CM g k
vk = pk +

k =

. . . . . . . . . . . . . . . . . . . . . . . gradient (A-3)

. . . . . . . . . . . . . . . . . . . . . . preconditioned (A-4)

pkT ( g k g k 1 )
vk 1
pkT g k

vkT g k
~

. . . . . . . conjugate gradient (A-5)

. . . . . . . . . optimal step-length (A-6)

vkT Avk + vkT CM1 vk

September 2001 SPE Journal

H
H
mk +1 = mk k vk

k = k + 1.

. . . . . . . . . . update model parameters (A-7)

In theory, the first term of the right hand side of Eq. (B-2) can be
calculated

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .(A-8)

2 J0
J 0
=

. . . . . . . . . . . . . . . . . . . . . . . . . . (B-3)
2

Because computing the inverse matrix of the covariance matrix CM


is costly, in this work we modify the preconditioned conjugate gradient method. Instead of computing the inverse matrix CM , the nondiagonal elements are set to zero when CM-1 is required. Let

Here, the optimal step length for the MAP estimates of the porosity
is as follows
~

CM = diag 1/ 12 1/ 22 L 1/ M2 .

. . . . . . . . . . . . . (A-9)

and the inverse matrix for CM can be obtained readily.


Appendix B
The proposed iterative scheme requires generating the secondorder derivatives of the flow equations with respect to permeability and porosity. However, the second-order partial derivatives of
the flow equations with respect to porosity do not exist. A transformation must be made.
If we define

T = 12 22 L M2 , . . . . . . . . . . . . . . . . . . . . . . . (B-1)
the second-order derivatives of objective function with respect to
j can be estimated
2 J 2 J0
=
2 2
T

n
X l n +1

l


l =o , w n = 0

N 1

X n +1 T

n +1
l

l .

bkT Abk + bkT CM1 bk = bkT Abk + vkT CM1 vk ,

. . . . . . . . . . . . . (B-4)

where
b=

T
T

n
n +1
X l n +1 X l n +1

l l .
l = o , w n = 0

N 1

SI Metric
ft
cp
psi

. . . . . . . . (B-5)

Conversion Factors
3.048*
E - 01 = m
1.0*
E - 03 = Pas
6.894 757
E + 00 = kPa

*Conversion factor is exact.

SPEJ

Zhan Wu is a postdoctoral research associate in the Petroleum


Engineering Dept. of Texas A&M U. e-mail: zhan-wu@
tamu.edu. His research interests include reservoir simulation,
reservoir modeling, and reservoir engineering. He holds an MS
degree from the Southwest Petroleum Inst. in China and a PhD
degree from the U. of Tulsa, both in petroleum engineering.

. . . . . . . . . . . . . . . . . . . . . . (B-2)

September 2001 SPE Journal

351

Вам также может понравиться