Sunteți pe pagina 1din 6

SPE 66350 Prediction of Oil Production With Confidence Intervals

James Glimm, SPE, SUNY at Stony Brook and Brookhaven National Laboratory; Shuling Hou, SPE, Los Alamos National Laboratory; Yoon-ha Lee, SUNY at Stony Brook; David Sharp, SPE, Los Alamos National Laboratory; and Kenny Ye, SUNY at Stony Brook. ensemble together with error probabilities for the forward solution. Results of the prediction methodology will be described, based on simulated geologies and simulated reservoir flow production rates. Efficient scaleup allows a sizable number of geologies to be considered. The Bayesian framework incorporates prior knowledge (for example from geostatistics or seismic data) into the prediction. We show that a history match to past production rates improves prediction significantly. The plan of this paper is to pick one fine grid reservoir from an ensemble and regard its solution as a stand in for production data. Other reservoirs in the ensemble are evaluated on the basis of the quality of their match to this data. They are upscaled, simulated on a coarse grid, and the upscaled solution is compared to production history from the data. Probability of mismatch between the coarse grid solution and the data weights each realization in a balanced manner according to (a) its prior probability and (b) the quality of its match to data. We thus define a posterior probability on the ensemble, which is used for prediction. Uncertainty in the prediction has two sources: uncertainty in the geology, or history match, as discussed above, and uncertainty in the forward simulation, also conducted on coarse grids. The total uncertainty receives contributions from these two sources, and its analysis leads to confidence intervals for prediction. The intended application of this prediction methodology is to guide reservoir development choices. For this purpose, simulation of an ensemble of reservoir scenarios is important to explore unknown geological possiblities. Statistical methods are important to assess the ensemble of outcomes. The methods are intended for use by reservoir managers and engineers. For this purpose, the methods will need to be augmented by inclusion of factors omitted from the present study. The significance of our methods is their ability to predict the risk, or uncertainty associated with production rate forecasts, and not just the production rates themselves. The latter feature of this method, which is not standard, is very useful for evaluation of decision alternatives. Stochastic History Matching Problem Formulation. Stochastic history matching is based on an ensemble of geological realizations. To simplify this study, we fix the geologic model aside from the

This paper was prepared for presentation at the SPE Reservoir Simulation Symposium held in Houston, Texas, 1114 February 2001. This paper was selected for presentation by an SPE Program Committee following review of information contained in an abstract submitted by the author(s). Contents of the paper, as presented, have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material, as presented, does not necessarily reflect any position of the Society of Petroleum Engineers, its officers, or members. Papers presented at SPE meetings are subject to publication review by Editorial Committees of the Society of Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper for commercial purposes without the written consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972952-9435.

Abstract We present a prediction methodology for reservoir oil production rates which assesses uncertainty and yields confidence intervals associated with its prediction. The methodology combines new developments in the traditional areas of upscaling and history matching with a new theory for numerical solution errors and with Bayesian inference. We present recent results of coworkers and ourselves. Introduction A remarkable development in upscaling1, 2 allows reduction in computational work by factors of more than 10,000 compared to simulations using detailed geological models, while preserving good fidelity to the oil cut curves generated from solutions of the highly detailed geologies. In common engineering practice, the detailed geology models are too expensive for routine simulation. This is especially the case if an ensemble of realizations of the reservoir is to be explored. The ensemble allows consideration of distinct geological scenarios, an issue of greater importance in many cases than errors associated with upscaling of detailed geology to obtain a coarse grid solution. Upscaling allows rapid solutions and is a key to good history matching. We formulate history matching probabilistically to allow quantitative estimates of prediction uncertainty3, 4. A probability model is constructed for numerical solution errors. It links the history match to prediction with confidence intervals. The error analysis establishes the accuracy of fit to be demanded by the history match. It defines a Bayesian posterior probability for the unknown geology. Thus history matching defines a revised ensemble of geologies, with revised probabilities or weights. Prediction is based on the forward solution, averaged with these weights. Confidence intervals are also defined by the probability weights for the


SPE 66350

permeability field, which is taken to be a random variable simple form of the Darcy and Buckley-Leverett equations5

According to Bayes theorem, the posterior probability for the geology defined by the permeability realization K is

v = K p = 0 ; v = 0 , .(1) s + v f (s ) = 0 , ....(2) t where is a relative mobility, K the absolute permeability, v velocity, p pressure, s the water saturation and f the fractional flow flux. We consider
these equations in a two dimensional (reservoir cross section) geometry, 0 x 1 , 0 z 1 in dimensionless units. Assume no flow across the boundaries z = 0,1 and a constant pressure drop across the boundaries x = 0,1 . The absolute permeability K is spatially variable, with an assumed log normal distribution. We characterize the covariance ln (K ) by correlation lengths l z = 1 / 50 and

p( K O) =

p (O K ) p ( K )

where p( K ) is the prior probability for the realization K. The prior probability is defined, for example, by methods of geostatistics 6, 7, 8, 9, 10, and in the present context it is defined by the above mixture of Gaussians with specified correlation lengths. In the absence of errors, there would be no mismatch, and we could accept geology K j as a history match only if

p (O K ) p ( K ) dK


c j (t ) fi0 (t ) . This is of course unrealistic, as errors do

occur. Since p O | K j

) assumes

K j = K i0 is exact, the

mismatch is assumed to be due to an error in determining c j . We write e j = f j c j as the error. Measurement errors also contribute to the mismatch likelihood, but for simplicity we concentrate on scale up and numerical solution errors only. Thus, p O | K j is the probability of

Thus, ln (K ) is actually a Gaussian mixture, and is not Gaussian. This distribution for K is called the prior distribution. Each realization is a specific choice of K . We consider an ensemble defined by 500 realizations of K , 100 for each of the five correlation lengths, selected according to the above Gaussian distribution. Each K is specified on a 100 x 100 grid (the fine grid). K and the fractional flow functions f are then upscaled to grids at the levels 5 x 5, 10 x 10, and 20 x 20. Each of the coarse grid upscaled reservoirs is also solved, in all cases for up to 1.4 pore volumes of injected fluid (1.4 PVI). We select one of the geologies, K i0 , as representing the

lx =

(0.2, 0.4 , 0.6, 0.8, 1.0) .

the error e j . A Probability Model for Solution Errors. To characterize the errors statistically, we introduce the mean

e (t ) = C s (s , t ) =

1 N

N j =1

e j (t ) , (4)

and the sample covariance

exact but unknown reservoir. We observe the oil cut f i0 generated by the fine grid solution for times

1 N (e j (t ) e (t ))(e j (s ) e (s )) , ...(5) N 1 j =1 s s The precision matrix is the operator inverse to C . We make the approximation e = 0 . Consider the expectation s (e, e) = e (s ) s (s , t ) e (t ) ds dt , .(6)
o t1

0 t t0 (PVI). This data represents past, historical data, and using it, we seek to predict production for t0 t t1 =
1.4 PVI, i.e. into the future. The solution is (a) history matching, to select a revised ensemble of geologies, which reflect agreement with history data, and (b) forward simulation, averaged over the revised (posterior) ensemble, to predict the future production. The Bayesian Framework. In the Bayesian framework, the prediction problem is solved by assigning a probability, or likelihood to any degree of mismatch between the coarse grid oil cut c j (t ) and the observed history

In Gaussian statistics, 1 2 (e, e) is proportional to the log probability of the error e . We also introduce the discrepancy dij (t ) = f i (t ) c j (t ) , .(7)

as the difference between the fine grid solution of one reservoir and the coarse grid solution of another. For i = j , we have dii = ei is an error. Bona fide discrepancies,

dij (i j ) are typically systematically

This fact allows us to

larger than errors e j = d jj .

O = f i0 (t ), 0 t t0 , where f i0 (t ) is the oil cut for the

reservoir K i0 computed on the fine grid (and the fine grid is conceptually considered to be exact). The probability or likelihood of the observation given the geology K is denoted p(O | K ) . A mismatch could arise due to measurement errors, or as we consider here, due to use of a coarse grid in a simulation analysis.

differentiate between errors (good matches to history) and discrepancies (poor match to history). See Fig. 1. The discrepancies play a critical role in evaluation of the Bayes likelihood p (O K ) . This likelihood assumes that the geology K has been chosen exactly. Hence, under this assumption, the discrepancy is an error, and the negative probability of the likelihood is ( d , d ) defined by (6). Thus,

SPE 66350



1 ln p (O K ) s (d , d ) , .(8) 2 d = d ji 0 if K = K j .

reduced, or history matched ensemble Ri those realizations 0 for which

f j ( s ) fi0 ( s )

, 0 s t0 , ..(9)

Here, t0 is the present time, and the tolerance is chosen subjectively, to make the history matched ensemble small, but not too small. For a small window size , this method requires a large starting ensemble. Oil cut curves for the

Fig. 1 Typical errors (lower, solid curves) and discrepancies (upper, dashed curves), plotted vs. PVI. The two families of curves are clearly distinguishable .

Model Reduction. The precision matrix and its inverse,


the sample coveriance C , are defined nonparametrically in terms of data. The data is a study of errors resulting from upscaling of a fine grid geology specification. According to well-established principles of statistics, the model or

C s for solution errors should be limited in complexity by

the amount of data used to define it. For this reason, we project the errors e to a p dimensional space, e p = p e , and equivalently, we truncate and C to obtain p p projected precision and covariance matrices
s s s sp = p s p and C p = p Cs p .

The general
Fig. 2 Top: Oil cut curves for the complete ensemble. Bottom: Oil cut curves for the reduced ensemble. The reduced ensemble is defined by (9) with

principles governing data smoothing state that p should be neither too small, to avoid bias, nor too large, to avoid dispersion, also known as overfitting of data. Conceptually, errors are averaged into p bins along the PVI axis, although better results can be obtained through higher order methods using finite element spaces4, 12. Prediction Window Based Prediction. We show a simple method of prediction, based on a window to accept or reject candidate reservoirs in the ensemble as being acceptable or unacceptable matches to history data. Window based prediction assumes a large number of fine grid solutions, and thus is not practical, but it serves to introduce the ideas. If the window is narrow, window based prediction is also more accurate, and thus by comparison to other prediction methods, it indicates the amount of information lost in various practical, but approximate prediction methods. We introduce a tolerance, or window size, and accept into a

= 0.08


f i0 (t0 ) = 0.8 .

unconstrained and constrained ensembles are shown in Fig. 2. To realize window based prediction within a Bayesian framework, we define a likelihood function p (O K ) to be 1 if the realization gives an oil cut lying within the window and 0 otherwise. A Metric for Prediction. We evaluate prediction by the percent reduction in error, relative to the base case of prior prediction. For any prediction p ( s ) of the future oil cut, we define the prediction error as

PEi0 = tt10 p ( s ) f i0 ( s ) ds , ...(10)

The choice


SPE 66350

p prior ( s ) =

1 N j = 1 f j ( s ) , ..(11) N

Table 2a. Bayesian Prediction Error Reduction (per cent). Present Oil Cut = 0.8
0.2 0.4 0.6 0.8 1.0 Mean

defines the fine grid prior prediction; similarly we define the coarse grid prior prediction with c j replacing f j in (11). The prior prediction error PE prior , io is defined by (10) with (11) substituting for p ( s ) . We thus define the per cent error reduction,

5x5 21 11 10 11 13 13

Level of Scale Up 10 x 10 40 09 08 11 22 18

20 x 20 45 14 07 08 15 18

ER = ( 1

PEio iN 0 =1 PE prior , io iN 0 =1

)100 , .(12)

Table 2b. Bayesian Prediction Error Reduction (per cent). Present Oil Cut = 0.6
0.2 0.4 0.6 0.8 1.0 Mean

To assess the window based prediction scheme, we define the restricted ensemble Ri0 as in (9), and

5x5 29 18 23 21 23 23

Level of Scale Up 10 x 10 50 25 28 28 33 32

20 x 20 55 22 33 27 45 36

p ( s) =

1 jR i f j ( s ) , .(13) 0 N i0

where N i0 is the number of elements of Ri0 . The resulting error reduction is presented in Table 1, first averaging over 100 choices of i0 of a specified correlation length, and then averaging these averages over all five correlation lengths. The three right columns in Table 1 specify initial times t0 for which the fine grid oil cut satisfies

In Fig. 3 we show the histogram for posterior predicted error prior predicted error for 10 x 10 scale up and a present oil cut of 0.6. The skewing of this distribution to negative

f i0 (t0 ) = 0.8, 0.6, 0.4 .

Table 1. Window Based Prediction Error Reduction (per cent)
0.2 0.4 0.6 0.8 1.0 Mean

0.8 53 33 37 42 43 42

Present Oil Cut 0.6 0.5 59 43 46 53 55 51 63 42 49 58 61 55

0.4 62 43 51 62 65 57

Bayesian Prediction. We use Bayes formula (3) with the likelihood (8) defined by the nonparametric error model (4), (5), (6). We define p using piecewise linear elements and a value of p =16 . We consider t0 , the present time, to correspond to an oil cut of 0.6. Bayes formula defines a weight, or probability, w j for each realization K j . The posterior prediction p ( s ) is then Fig. 3 Histogram of posterior predicted error prior predicted error. values represents the improvement (and sometimes a very significant improvement) of the posterior over the prior. Confidence Intervals. A confidence interval is a range of possible outcomes likely to include the true outcome. For example, a 5%-95% confidence interval is an interval of outcomes that will include the true outcome with 90% probability, but with a 5% chance that the outcome will lie outside the interval on either the low or the high side. To determine confidence intervals, we need to convolve two sources of uncertainty or error: the choice of the correct geology and the accuracy of the forward simulation, for

p( s ) = N j = 1 w j c j ( s ) , ..(14)
We use (11) with f j replaced by c j to define the prior prediction and (10), (12) to define per cent error reduction. The results are presented in Table 2 for three levels of scale up (5 x 5, 10 x 10, 20 x 20). Table 2a gives error reduction percentages for a present oil cut of 0.8. Table 2b assumes a present oil cut of 0.6. Results for a present oil cut of 0.4 (not presented here) show continued improvement.

SPE 66350


times t in the future, t0 t t1 , using a coarse grid with known error statistics. To assign probabilities to outcomes, we need to reassess our metric for prediction. We consider the total future oil production the predicted production

t1 t0

f i0 ( s )ds, compared to

Table 3. 5%-95% Confidence Prediction of Future Production, as Future Production. Comparison Standard Deviation and Mean as Future Production.

Intervals for a Per Cent of to Prediction Per Cent of

Mean 1 -4 -8 -14 -15 -8

t1 t0

p( s )ds. Thus, we remove the

absolute values in (10). In place of the mean prediction p ( s ) we need to examine the probability density function ( pdf ) for the corresponding random variable

0.2 0.4 0.6 0.8 1.0 Mean

Confidence Intervals [-25,9] [-27,12] [28,13] [-30,17] [-29,14] [-28,13]

Standard Deviation 6 11 20 28 25 18

ti t0

f j ( s ) ds =

ti t0

(c (s) + e (s) )ds , ..(15)

j j

Here the random variables are ( a) j , governed by the Bayes posterior distribution, and (b) e j , governed by both Bayes posterior and by the model for error statistics. Here we introduce an approximation, and evaluate e j in the prior distribution, i.e. to be independent of the observed oil cut

f i0 in the past, and independent of the random variable c j

We make this approximation because error statistics are expensive to generate, and in practice their customization to specific reservoirs may not be feasible. To find confidence intervals for (15), we sort values by magnitude, and exclude the top and bottom 5% in probability. The outer boundaries ci0 retained in this sort are the confidence intervals. The result still depends on the choice of exact geology i0 . To facilitate comparison of confidence bounds for distinct i0 , we non-dimensionalize the confidence intervals to obtain confidence intervals for relative error

Discussion The prediction error reduction is consistent with earlier results3, 4 based on a smaller sample size (50 vs. 500 realizations). The error reduction shows understandable trends. As the present time is moved later, more information is available to constrain the predictions, and the reduction of prediction error increases. (The error decreases). As the problem is solved more accurately (with less scale up), the error reduction also increases. The window based predictions using only fine grid solutions are consistently more accurate than the Bayesian posterior predictions using upscaled (approximate) solutions. The best predictions, and largest error reductions are obtained with the smallest correlation lengths, i.e. for reservoirs not dominated by narrow conduction bands. Such reservoirs show less variability for production and are thus easier to predict. We note a systematic tendency (bias) for the upscaled solutions to overpredict production, especially for the highly layered reservoirs (with long correlation length). Conclusions A systematic method is presented for assessing uncertainty and combining incomplete information. Scientifically based probabilities allow improved management of risk. The benefit of this technology is a rational, objective, and systematic method to combine all available partial knowledge, to formulate to make predictions, based on this knowledge, and to assess uncertainty. Key ingredients of this technology are: 1. Upscaling for rapid solutions 2. Error estimated to generate probabilities and likelihoods 3. Bayesian statistical inference With these methods, we achieve a reduction of prediction error by 30% in comparison to prior prediction, for a sample problem. We achieve 5%-95% relative percentage confidence intervals of [-28%, +13%]. Further studies are needed to assess various methods to describe uncertainty in prediction. We will also investigate trade offs between increased computational work (ensemble size, degree of scale up) and the narrowing of confidence intervals. Finally, it will be of interest to find the inherent limits, set by lack of knowledge, for accuracy of prediction. Nomenclature

ci tt1 p ( s ) ds 0

t1 t0

p (s ) ds

, (16)

The confidence intervals (16) have a weaker dependence on

i0 than do the ci . We average them. For comparison, we 0

report the mean (with respect to average over i0 ) non-dimensionalized prediction error of the

t1 t0


( s ) p ( s ) ds f i0 ( s ) ds
, .(17)

t1 t0

The mean refers to a systematic bias in the prediction methods. We also report the standard deviation for the prediction, defined relative to both the Bayes posterior and the i0 summations. See Table 3. The standard deviation contains information similar to that of the confidence intervals. For Gaussian errors, the 5%-95% confidence intervals are equal to 1.645 , i.e., a multiple 1.645 of the standard deviation. Here we consider 10 x 10 scale-up, p (binning dimension) = 16, present oil cut = 0.6.

Cs ci (s )

= =

sample covariance coarse grid oil cut for geology i


SPE 66350

dij ( s ) ei (s ) ER f i ( s) lx , lz Ki N O p (K ) p (O K ) p ( K O) p (s) PE PVI R s, t t0 t1 wj s p


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

discrepancy error error reduction (per cent) fine grid oil cut for geology i correlation lengths absolute permeability of geology

Wallstrom, T., Hou, S., Christie, M. A., Durlofsky, L. J. and Sharp, D. H. Accurate scale up of two phase flow using renormalization and nonuniform coarsening. Computational Geoscience, 3:69-87, (1999). 2. Wallstrom, T., Hou, S., Christie, M. A., Durlofsky, L. J. and Sharp, D. H. Application of a new two-phase upscaling technique to realistic reservoir cross sections. Proceedings of the SPE 15th Symposium on Reservoir Simulation, pages 451-462, (1999). SPE 51939. 3. Glimm, J., Hou, S., Kim, H., Sharp, D. and Ye, K. A probability model for errors in the numerical solutions of a partial differential equation. CFD Journal, 9, 2000. Proceedings of the 8th International Symposium on Computational Fluid Dynamics, Bremen, Germany, (1999). 4. Glimm, J., Hou, S., Kim, H., Lee, Y., Sharp, D. H., Ye, K. and Zhou, Q.: Risk Management for Petroleum Reservoir Production: A Simulation-Based Study of Prediction, submitted to J. Comp. Geosciences. 5. Peaceman, D. Fundamentals of Numerical Reservoir Simulation. Elsevier, Amsterdam-New York, (1977). 6. Haldorsen, H. and Lake, L.: A new approach to shale management in field scale models. In SPE Res. Eng. J., (1984) pages 447-452. 7. Journel, A. G. and Huijbregts, Ch. J. Mining Geostatistics. Academic Press, New York, (1978). 8. Lake, L. and Carroll, H. editors. Reservoir Characterization. Academic Press, New York, (1986). 9. Lake, L. and Carroll, H. and Wesson, T. editors. Reservoir Characterization II. Academic Press, New York, (1991). 10. Deusch, C. V. and Journel, A. G.: Geostatistical Software Library and Users Guide. Oxford University Press, Oxford, (1992). 11. Hrdle, W.: Smoothing techniques: with implementation in S. Springer-Verlag, New York, (1991). 12. Strang, G. and Fix, G.: An Analysis of the Finite Element Method. Prentice-Hall, Englewood Cliffs, (1973). 1.

number of realizations observation (of oil cut history) Bayes prior Bayes likelihood function Bayes posterior predicted oil cut prediction error pore volumes injected fluid restricted ensemble of reservoirs times (in PVI ) present time final time ( = 1.4 PVI ) posterior probability of realization K j = = = variance sample precision operator projection onto p dimensional binning subspace

i, j i0 p

= = =

indices for sample geologies geology taken as exact dimension of data subspace used for binning or error model reduction

Acknowledgments The work of James Glimm is supported in part by the NSF Grant DMS-9732876, the Army Research Office Grant DAAG559810313, the Department of Energy Grants DEFG02-98ER25363 and DE-FG02-90ER25084, and Los Alamos National Laboratory under Contract #C738100182X. Shuling Hou is supported by the Department of Energy under Contract W-7405-ENG-36. Yoon-ha Lee is supported by the Department of Energy grant DE-FG02-90ER25084. David Sharp is supported by the Department of Energy under Contract W-7405-ENG-36. Kenny Ye is supported in part by Los Alamos National Laboratory under Contract #C738100182X.