Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Reservoir Forecasting
C.S. Kabir and A. Chawathe
, ChevronTexaco, and S.D. Jenkins, A.J. Olayomi, C. Aigbe,
and D.B. Faparusi, Chevron Nigeria Ltd.
Summary the reservoir. In addition, the traditional tornado chart, used for
Limited and uncertain geologic and engineering data at the onset ranking the variables, does not provide any information pertaining
of any new field development are the bane of reservoir character- to statistical significance of the independent-variable effects on
ization and simulation. The problem stems from the uncertainty in the dependent variable.
various model-input variables, such as reservoir connectivity, fluid However, systematic approaches19 have emerged to account
viscosity, and endpoint saturations, to name a few. Given this for uncertainty associated with various input variables, on the basis
scenario, an ad hoc, one-factor-at-a-time approach to earth and of ED. For instance, Chewaroungroaj et al.5 demonstrated the use
flow-simulation modeling cannot possibly yield unbiased informa- of ED with a series of dimensionless variables to allow extension
tion for making objective business decisions. of the results to similar systems. On the other hand, Corre et al.7
This study presents three field cases in which both engineering used ED to integrate data from diverse sources, including seismic,
and earth-model variables were varied in a systematic way to in their effort to quantify uncertainty. Application of ED also has
assess reservoir performance by use of the experimental design been extended to history-matched reservoirs.8
(ED) approach. In ED, geologic (e.g., fluid contacts, net-to-gross pay thickness)
Results of the field cases show that well requirements (both and engineering (e.g., anisotropy, well count) variables are varied
producers and injectors) turned out to be fewer than anticipated. simultaneously, unlike the ad hoc, one-variable-at-a-time ap-
Equally important, one case study showed that laboratory mea- proach. The advantage of the ED approach is that it generates
surements could minimize uncertainty surrounding oil viscosity relatively unbiased p-10, p-50, and p-90 probabilistic estimates by
and endpoint saturations. At the same time, we learned that the capturing nonlinear interactions of variables. Full-factorial three-
preferred horizontal-well orientation was marginally superior to level designs, which account for nonlinearities, are currently rea-
vertical wells in light of high reservoir anisotropy. In another case, sonable with analytic simulations10 or when only a few (n<4)
stratigraphy, gas/oil contact (GOC), and aquifer strength became variables are involved. Realistically, with finite-difference simu-
the primary variables for the full-factorial design after the initial lations, one can afford to do only a subset of the simulations at a
screening. Here, we proved that the project could proceed because cost of some higher-order interactions between the variables. Here,
it met the minimum reserves criterion. Perhaps most importantly, we chose to rely on the statistical principle of sparsity; that is, the
all studies showed how to obtain unbiased information in far fewer main effects of a few variables cause most primary influences.
flow-simulation runs than one would do using an ad hoc approach. In this paper, we used a three-step procedure that involves
variable screening using a simple, linear design, followed by a
nonlinear analysis and the response surface approach. This meth-
Introduction
odology is outlined in detail in Friedmann et al.2 and is similar to
The availability of very limited data with large uncertainty pre- the one originally presented by Damsleth et al.1
sents significant challenges to any new field development. Stakes This paper demonstrates the use of ED for making rapid and
are high when deepwater prospects are evaluated. With advances unbiased business decisions and, therefore, differs from those
in 3D seismic, the reservoir surface may be mapped with a certain available in the literature. In particular, the paper stresses an op-
degree of confidence. However, a few exploratory wells cannot erating-company approach, vis-a` -vis developing new data-
provide detailed information about the reservoirs internal archi- gathering strategies and the like. We also discuss the limitations of
tecture, particularly with respect to flow barriers or baffles. In ED, especially in an earth-science context, wherein variable
short, we are confronted with large uncertainty in the reservoirs screening tends to be more qualitative than quantitative in nature.
flow and, sometimes, fluid properties.
Given limited and uncertain data, questions arise about how to Methodology
proceed with a field-development plan. Historically, we have used In essence, the proposed approach entails a three-step procedure
reservoir simulation and sensitivity analyses as tools for predicting using ED. First, we screen a large number of variables by using the
various scenarios, followed by economic analysis. But the ap- two-level (low and high) design of Plackett and Burman (PB).11
proach has been less than satisfactory because of the ad hoc nature This step is designed to identify the major variables, or heavy
of the exercise, meaning the need to change one variable at a time. hitters, influencing the dependent variable (recovery, for in-
In essence, this approach relies on setting one variable at the p-10 stance) by capturing only the linear effects. Rapid screening with
or p-90 level while keeping others at the p-50 level in a simulation a minimum number of flow simulations is the focal point here.
run. Subsequent ranking of the independent variables can conceiv- Second, the time-lapse recovery (or any other dependent vari-
ably be biased. Potentially, this bias stems from two sources. First, able) is used to rank the sensitivity of the major variables, identi-
simulations may not contain independent information in the result- fied in Step 1. The variables belonging to the 95% confidence
ant 2n+1 simulations, where n represents the number of variables. interval are selected for the three-level (low, mid, and high) D-
Second, the method relies on comparing solutions in which all but optimal or full-factorial design. Thereafter, flow simulations are
one variable are set at the p-50 level. Finally, when one uses p-50 run on the basis of this design. This step is designed to capture both
values in the conventional analysis, the p-50 outcome may be an linear and nonlinear effects in addition to interactions between
unrealistic expectation owing to nonlinear fluid-flow behavior in the variables.
Third, we generate a response surface with multivariate analy-
sis by fitting a polynomial, which then serves as a proxy to the
flow simulator. To fill in the information void generated by the few
Copyright 2004 Society of Petroleum Engineers
flow-simulation runs, Monte Carlo simulations are done with the
This paper (SPE 87643) was revised for publication from paper SPE 77564, first presented response surface to generate the cumulative distribution function
at the 2002 SPE Annual Technical Conference and Exhibition, San Antonio, Texas, 29
September2 October. Original manuscript received for review 13 February 2003. Revised
(CDF). Recovery factors corresponding to different probabilities
manuscript received 6 August 2003. Paper peer approved 12 December 2003. are then derived from the CDF.
Fig. 1Pareto chart showing recovery after 5 years. Fig. 2Pareto chart showing recovery after 15 years.
Fig. 12Relative contributions of variables to the recovery Fig. 13Relative contributions of variables to the recovery
factor at 5 years, Case 2. factor at 19 years.
at the time of recent budget allocation. This project will get a a given OOIP. Of course, major changes in reservoir connectivity
favorable management response when the risks are mitigated by may have a significant impact on a development plan.
codevelopment of gas reserves in this field. In the first two case studies, we presumed a single geostatistical
We point out that risk mitigation is an important element in the realization to be valid for answering the questions that we posed
proposed approach by gathering additional data through new seis- (i.e., how many producers and injectors were needed for field
mic acquisition, new drill, and the like. Continuous updating of the development). Experiences suggest that different geostatistical re-
model and recomputation of CDF must follow for reassessment of alizations do not necessarily yield different solutions when busi-
the projects economic value at a given point in time. ness decisions are made on the basis of recovery. This point was
made when the two geologic models, quick-look and detailed,
Discussion yielded very similar recoveries in Case 2. However, Case 3 indi-
Although we studied the uncertainty of various variables, the issue cates that major geologic uncertainties, such as stratigraphy, will
of optimal well location was not explored explicitly. Rather, we impact recovery. But we caution that such variables are extremely
relied on a deterministic approach for ascertaining well locations, difficult to quantify in continuous probabilistic terms, thereby
independent of the ED approach. We recognize that there are ways questioning the estimation of low, mid, and high identities.
to optimize well locations en route to maximizing an objective Some geological variables, such as structure, are even more
function, such as net present value.14,15 Gu yagu ler and Horne14 difficult to quantify because of their inherent spatial nature. Avail-
extended the well-optimization strategy with a hybrid genetic al- able algorithms can produce a CDF of structures (and/or geologic
gorithm (GA)-polytope method. Unfortunately, the current well- models), which may be ranked on the basis of OOIP. However, we
optimization strategies1416 are not well suited for multiple geo- perceive this approach to be of limited value because a true un-
logical realizations and complex well combinations. However, certainty measure is incomplete without incorporating a develop-
given the size of the prospects and the degree of heterogeneity, we ment plan, which leads us back to the chicken-and-egg situation.
think that the ED approach used here suffices. Experiences under- We also think that the presence of yet-to-be-established flow
score the need to find ways to incorporate the development plan barriers will have the most impact on reservoir fluid flow and,
early in the uncertainty-estimation process. However, this notion therefore, on recoveries. To improve the odds, we recommend
sometimes leads to the chicken-and-egg situation; that is, a devel- updating these relatively inexpensive studies as additional data are
opment plan is based on a specific geologic model, but what is a gathered with the drilling of development wells. In this way, one
p-10 (or p-50, or p-90) geologic model without a development can minimize uncertainty in forecasting the performance of
plan? One recent study17 suggests that the well count, a measure of new fields.
development plan, does not change with rock properties alone for Finally, there is the issue of uncertain variables vs. engineer-
ing sensitivities. Some variables can be controlled, such as the well
count. On the other hand, permeability cannot be controlled and is
considered to be a pure uncertainty. In an ideal scenario, we rec-
ommend that the variables be split into two separate designsone
for the uncontrollable variables, and the other for the control-
lable variety. In general, experience indicates that engineering
variables (controllable) tend to dominate the reservoir performance
and, therefore, if coupled with the uncontrollable subsurface
variables, might overshadow the pure subsurface uncertainty.