Professional Documents
Culture Documents
Errors in History Matching PDF
Errors in History Matching PDF
Summary forecast. There have been some studies that attempt to tackle the
The usual procedure in history matching is to adopt a Bayesian problem of nonuniqueness of the solution, for example Refs. 1, 6,
approach with an objective function that is assumed to have a and 53 through 55. Ref. 53 presents a field story in which the risk
single simple minimum at the “correct” model. In this paper, we arising from such nonuniqueness is highlighted. In Ref. 6, an ap-
use a simple cross-sectional model of a reservoir to show that this proach is proposed that couples a chaotic sampling of parameter
may not be the case. The model has three unknown parameters: space with a local minimization technique. Through the evolution
high and low permeabilities and the throw of a fault. We generate of a nonlinear dynamical system, several points are successively
a large number of realizations of the reservoir and choose one of used as initial guesses for a local gradient-based optimizer. This
them as a base case. Using the weighted sum of squares for the provides a series of alternative matched models with different
objective function, we find both the best production- and best production forecasts that improve the understanding of the pos-
parameter-matched models. The results show that a good fit for the sible reservoir behaviors. In Ref. 1, the application of a global
production data does not necessarily have a good estimation for the optimization algorithm called the tunneling method is presented
parameters of the reservoir, and therefore it may lead to a bad for two test cases in which a series of minima was found. In Ref.
forecast for the performance of the reservoir. We discuss the idea 54, there is a study on the parameter estimation in the oil industry
that the “true” model (represented here by the base case) is not carried out by applying a modified genetic algorithm on a simple
necessarily the most likely to be obtained using conventional his- reservoir model. It demonstrated that it is difficult to obtain a
tory-matching methods. history match using simple optimization methods. The simple
model used in Ref. 54 was a 2D cross section of a reservoir with
a sequence of alternating good- and poor-quality sands and a
Introduction
simple fault. The three unknown parameters were the high and low
Reservoir history matching is a difficult inverse problem arising in permeabilities of the sands, and the throw of the fault. In a later
the petroleum industry. The aim of history matching is to find a work by J.N. Carter,55 there is an attempt to capture the effects of
model such that the difference between the performance of the modeling errors in inverse problems using Bayesian statistics. He
model and the history of a reservoir is minimized. Traditionally, then used a similar cross-sectional model of a reservoir to test the
this is done by hand. But the task of varying the parameters of a results. It appeared that the proposed error model yields a multi-
reservoir description by hand until a satisfactory match is obtained modal objective function that leads to multiple acceptable solu-
is extremely onerous and time-consuming. Therefore, gradient- tions. The approach used in Carter55 was to generate a large num-
based optimization techniques are increasingly adopted by the oil ber of models, which could be searched to find the best match
industry for computer-aided history matching. This is because of according to the criteria chosen.
the great time-saving benefits they can offer over conventional In this paper, we use the same cross-sectional model of a res-
trial-and-error approaches. Starting with an initial reservoir de- ervoir as in Refs. 54 and 55, and we also use an approach similar
scription, these optimization techniques automatically vary reser- to the one in Carter55 to generate a large number of realizations of
voir parameters until stopping criteria are achieved and a history the reservoir. We then choose one of these models as a base case,
match of field performance is obtained. This mechanistic activity and use the sum of squares as the measure of fit between the
is referred to as “automatic history matching.” In Refs. 1 through performance of the base case and each realization. We also find the
18, there are samples of studies that use the method of gradients for “distance” between the two models on the parameter space. By
automatic history matching in order to find the best-matched obtaining both the best production- and the best parameter-
model of a reservoir. Many other methods have been studied and matched models, we then study the history matching and also the
applied to the history-matching problem. Examples include opti- prediction of the both models. The results show that a model with
mal control theory,19–22 stochastic modeling techniques,23–26 sen- a good history matching could have parameters quite different
sitivity analysis techniques, 27–29 and gradual deformation from those of the “true” case, and it may give a bad forecast. To
method.31 In Refs. 31 through 45, some other studies on history show that the results are generic, a general study of the forecast
matching are listed. The use of 3D streamline paths to assist in then is carried out that confirms the robustness of the results. For
history matching has become increasingly common, because of the rest of the paper, firstly in the next section, the reservoir model
their speed compared to that of the conventional methods. Samples is described and objective functions are defined. Then, a base case
of such studies are given in Refs. 46 through 52. These techniques is chosen in the best-models section, and the best production- and
are designated as “assisted history matching” to distinguish them parameter-matched models are obtained. Then, in the following sec-
from automated history-matching techniques. tion, the prediction of the best models is assessed, and then a gener-
In most of the studies above, independent of the method used alized discussion of the forecast is given, followed by conclusions.
for the history matching, there is usually an assumption that there
exists a simple unique solution at the “correct” model. They there- Reservoir Model and Objective Function
fore neglect the inherent nonuniqueness of the solution of the
We use a simple 2D cross-sectional model of a layered reservoir,
underlying inverse problem. This, consequently, leads to the as-
as shown in Fig. 1.57 The model is 2D, with no-flow boundary
sumption that a good history-matched model is a good represen-
conditions. Water is injected on the left side, and a production well
tation of the reservoir and therefore gives a good forecast.
is located on the right. There is a single, normal fault at the mid-
In this paper, we want to challenge these assumptions and show
point between the two wells. The reservoir contains six alternating
that the problem may have multiple solutions, and subsequently, a
layers of good- and poor-quality sands with high and low perme-
good history-matched model might have geological properties
abilities, respectively. The porosities of the good- and poor-quality
quite far from those of the “truth” and therefore could lead to a bad
sands are 0.30 and 0.15, respectively. We also assume that the
three good-quality layers have identical properties, and the three
poor-quality layers have a different set of identical properties. The
Copyright © 2004 Society of Petroleum Engineers thickness of the layers has an arithmetic progression, with the top
Original SPE manuscript received for review 6 August 2003. Revised manuscript received
layer having a thickness of 12.5 ft, the bottom layer a thickness of
21 April 2004. Paper (SPE 86883) peer approved 28 May 2004. 7.5 ft, and with a total thickness of 60 ft. The width of the model
is 1,000 ft. The viscosity used for water was 0.31 cp; for oil, w = 0.03Qwb共h0,kp0,kg0,ti兲, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (2)
between 0.51 and 0.74 cp within the pressures of 4,041.7 and
9,014.7 psi. Power-law forms are used for the water and oil relative o = 0.03Qob共h0,kp0,kg0,ti兲. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (3)
permeabilities as functions of the water and oil saturations, respec-
tively. To produce a set of measurements, we therefore need three We normalize ⌬m by dividing it by the number of the mea-
properties: the size of the fault throw h, the permeability for the surements, n.
good-quality sand kg, and the permeability of the poor-quality sand We also define a measure of “distance” between the base case
kp. The three parameters are selected independently from uniform and each realization on the parameter space as ⌬p,
distributions with ranges: fault throw, h ∈ [0,60], good-quality
permeability, kg ∈ [100,200], and poor-quality permeability, kp ∈ 3
1 共Pi − Pbi兲2
[0,50]. The simulation model is 100×12 gridblocks, with each
geological layer divided into two simulation layers of equal thick-
⌬p = 兺
3 i=1 2pi2
, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (4)
再冋 册
We choose a point in the parameter space to be the base case;
2
h0⳱10.4 ft, kg0⳱131.6 md, and kp0⳱1.3 md. The production
n
Qw共h,kp,kg,ti兲 − Qwb共h0,kp0,kg0,ti兲
兺
1
= rates obtained are shown in Fig. 2. We then randomly sample
n i=1 2w2 共ti兲
another 159,645 realizations of the reservoir from the parameter
冋 册冎
2
space and run the simulator to obtain the production data for each
Qo共h,kp,kg,ti兲 − Qob共h0,kp0,kg0,ti兲 sample model. Obtaining objective function ⌬m gives a measure of
+ . . . . . . . . . . . . . . . . . (1) fit between the production data of the base case and each realiza-
22o 共ti兲
tion at the history-matching period, which we assume to be 3
Here, index m stands for the match of the history. Qw and years, n⳱36 months. Also calculating ⌬p from Eq. 4 gives a
Qwb⳱the water production rates for each realization and the base measure of distance between the base case and each realization on
case, respectively, at each measurement time ti. Similarly, Qo and the parameter space.
Qob⳱the oil production rates for the same models. The subscript
b corresponds to the base case, and n⳱the number of the mea- Best Production-Matched Model. Here, we first look for one
surements during the history-matching period. o and o⳱the single model which has the best match to the production data and
standard deviations of the water and oil production rates, respec- call it the best history-matched model. This is a representation of
tively, of the base case. In this paper, we consider the standard common exercise that is usually done in the industry. The model
deviations as weighting factors for normalization purposes. We for which ⌬m is minimum is the one with the best production
assume the weighting factor to be 3% of the measuring value match. In our case, the parameters for such a model are h⳱33.1 ft,
(either water or oil production), which gives kg⳱135.9 md, and kp⳱2.62 md, which gives ⌬m⳱0.118. In
Fig. 3, we plot the oil/water production rates for both the base case of the base case, and kp is almost twice as large as in the base case.
and the best production-matched model for 4 years from the start These suggest that a model with a good fit for the production data
of the production. The vertical dotted line corresponds to the end of a reservoir does not necessarily have a good estimation for the
of the history matching (3 years). We see that there is a very good parameters of the reservoir.
fit between the production profiles of the two models during the
history-matching period. For these two models, we also calculate Best Parameter-Matched Model. Alternatively, we may look for
⌬p for a measure of fit between their parameters, which gives a the model with the best parameter match. The realization with the
very large value, ⌬p⳱1,066. This value indicates there is a bad best parameter fit to the base case is the one with the minimum
match between the parameters of the two models. For the best value of ⌬p. We obtain such a model with parameters of h⳱11.15
model, kg is quite well estimated, h is almost three times the throw ft, kg⳱134.8 md, and kp⳱1.35 md, which gives ⌬p⳱1.23. For this
Fig. 3—Oil/water production rates vs. time for the base case (solid line) and the best production-matched model (dashed line) for
4 years from the start of production. The vertical dotted line shows the end of the history matching (3 years).
Fig. 4—Oil/water production rates vs. time for the base case (solid line) and the best parameter-matched model (dashed line) for
4 years from the start of production. The vertical dotted line shows the end of the history matching (3 years).
Fig. 6—Saturation map of the best production-matched model at the end of history matching.
Fig. 7—Saturation map for the best parameter-matched model at the end of history matching.
Fig. 8—Depth vs. saturation at the production well for the “true” model 1 year after the end of history matching.
Also, to see any possible relation between the goodness of the consider the two distinct sets of models with ⌬m<1, which have
forecast for every model with the closeness of the parameters of relatively good history match: those with good prediction, ⌬f<20,
that model to the base case, we plot ⌬f vs. ⌬p. These graphs are and those with poor forecast, ⌬f>20. For each set, we fit a Gaussian
plotted in Figs. 11 and 12. There seems to be a broad correlation distribution to the likelihood of ⌬m over the parameter space using
between the forecast and both the history of production and the the least-squares method. As the prior probability is uniform, the
parameters of the models. This is such that very large values of likelihoods represent the posterior probabilities for the two modes.
⌬m , which indicate very bad history-matched models, correspond We then find the values for the mean and standard deviation of
to a very bad forecast with high values of ⌬f. The same follows for each parameter for both Gaussian distributions. The results are as
the relationship between ⌬f and ⌬p. However, we are interested in follows. For the cluster with ⌬m<1 and ⌬f<20, the measured means
studying the forecasts of the realizations with productions or pa- are 1⳱5.20, 2⳱127.0, and 3⳱2.74, and the standard devia-
rameters relatively close to those of the “true” case. We therefore tions are 1⳱0.30, 2⳱0.37, and 3⳱0.52. For the models with
consider models in Figs. 11 and 12 with small values of ⌬m and ⌬p. ⌬m<1 and ⌬f 20, the fitted Gaussian distribution gives 1⳱39.4,
Figs. 13 and 14 represent these graphs for when ⌬m<5 and ⌬p<10, 2 ⳱136.7, and 3 ⳱2.53, and the standard deviations are
respectively. The points at the origin correspond to the base case 1⳱3.04, 2⳱1.88, and 3⳱0.004. In these values, the indices 1,
with ⌬m⳱⌬p⳱⌬f⳱0. The best production- and parameter- 2, and 3 correspond to the parameters h, kg, and kp, respectively.
matched models are indicated by arrows with, respectively, lowest
⌬m and ⌬p in Figs. 13 and 14. The corresponding ⌬f for these
Conclusions
models are 32.7 and 7.87, respectively, for the former and the latter
models. These show a better prediction for the best parameter- We studied a simple 2D cross-sectional model of a reservoir with
matched model, and a worse forecast for the best history-matched alternating layers of good- and poor-quality sands with high and
model, which are similar to the results we obtained in the section low permeabilities, and a simple fault in the middle. The model
on best-model prediction. However, the plots in Figs. 13 and 14 in then had three unknown parameters: the high and low permeabili-
general do not seem to indicate any systematic correlation between ties, kg and kp, respectively, and the throw of the fault, h. We ran
the goodness of the forecast ⌬f and that of the history match, ⌬m, a simulator for several realizations of the reservoir and obtained
or parameter match, ⌬p. Well history-matched models (with small the production data. Then, we chose a model as a base case and
⌬m) or models with parameters close to those of the base case assumed the end of the history matching to be shortly after water
(with small ⌬p) both might lead to bad forecasts with relatively breakthrough in the base case (3 years). As a goodness-of-time
large values of ⌬f. We also see that a good forecast with small ⌬f measure we used a weighted sum of squares of the difference
might correspond to small or large values of ⌬m or ⌬p. This seems between the production of each realization and the base case until
to suggest that models with bad history matching or those with the end of the history matching, ⌬m. We also defined a measure of
parameters far away from the “true” case still might result in “distance” on the parameter space between the base case and each
relatively good prediction of the production in a reservoir. realization, ⌬p. We then found both the best production-matched
The other point in Fig. 13 is the presence of two distinct clus- and the best parameter-matched models. The results showed that
ters: one with models of good prediction (low values of ⌬f) and the the best production-matched model does not necessarily have a
other with models of bad prediction (high values of ⌬f). Here, we good fit for the parameters of the reservoir. On the other hand, a
Fig. 10—Depth vs. saturation at the production well for the best parameter-matched model 1 year after the end of history matching.
model with its parameters close to those of the base case might not a saturation profile very similar to that of the base case at the end
have a good match to the production data. of the forecast. This suggests that a good history-matched model
We then looked at the forecast of the reservoir for both best with productions close to those of the “true” case does not neces-
models one year after the end of the history matching. The quan- sarily have a good forecast, and therefore might not be a good
tities we considered were the production data and the water- representative of the “truth.” However, it seems that the model
saturation profile at the production well. The production data of the with parameters close to those of the base case might lead to a
best history-matched model and of the base case seem to diverge better forecast for the “true” case. To show the robust nature of the
during the forecast period. The two models show very different results above, we then studied the forecast on a more general basis.
saturation profiles at the end of the forecast. The production data We therefore introduced a goodness-of-fit measure, ⌬f, between
of the best parameter-matched model, however, seem to converge the production data of the base case and each realization from the
toward those of the base case in the prediction period, which gives end of the history matching until the end of the forecast. By plot-
ting ⌬f vs. ⌬m and also ⌬p, we did not observe any general sys- We would also like to mention that the problem we looked at in
tematic correlation between the goodness of the forecast with ei- this paper is well-posed. In a typical history-matching exercise, the
ther production match or parameter match. Well-history-matched number of data values available are typically less than the number
models or models with parameters close to those of the base case of unknown parameters in the reservoir, so it is an ill-posed prob-
could both lead to relatively bad forecasts. Alternatively, models lem and leads to multiple solutions. However, in our case, there are
with bad history matching or those with parameters far away from only three possible variables with 108 measurement data with no
the “true” case might still result in a relatively good forecast for the modeling errors, but still we run into errors in history matching.
reservoir. To show that the results obtained in this paper are more Therefore, the results obtained here are not caused by any ill-
generic, we have carried out a further study in Ref. 56, which posedness of the problem. Consequently, using more data might
investigates the errors both in history matching and in prediction change the so-called best history-matched model, but the prin-
by systematically varying each parameter in the reservoir model. ciples of the occurrence of errors will remain unchanged.