History Match and Associated Forecast Uncertainty Analysis—Practical Approaches Using Cluster Computing

This paper presents practical approaches to deal with the complex problem of the uncertainty assessment in the performance forecast using reservoir simulation models with extensive production history. The complexity and difficulty of this type of problem arises mainly from the necessity of finding a large number of simulation models that are consistent not only with the geological data but also with the observed production history. In simpler terms this means finding an appropriate number of multiple solutions to the history match problem that can be used to estimate uncertainty in the forecasts. The rigorous solution to this kind of problem involves the application of methods based on Monte Carlo simulation; but they are not routinely applied because of the computational cost associated to the necessary large number of simulations for real field problems. Advances in computing technology in recent years, especially in the areas of CPU speed and of high performance computing affordability with medium to large CPU clusters, indicate that now is, probably the appropriate time to explore and revisit the practical aspects of performing a more comprehensive history match and forecast uncertainty analysis with Monte Carlo simulation methods. The approaches presented in this work take advantage of the availability of a medium size 256 CPU Linux cluster that allowed the coupling of distributed high performance computing with efficient sampling techniques to solve the history match and the associated forecast uncertainty problem under a probabilistic inverse problem framework, and to present the results of both history match and forecast in the form of probability density functions (PDF). Prior probabilistic model information is incorporated in the process. The tests performed with data from a real field indicated that our approaches provide one practical way to address, more comprehensively than current existing approaches, the non-uniqueness issue of the history matching problem and the associated uncertainties in performance forecasts in real fields. Since the results are accomplished in a very short time, significant changes in reservoir management paradigms may result. Introduction History matching is an inverse problem to calibrate reservoir simulation models to the observed production history, and it is a critical and necessary step in optimizing reservoir management decisions associated to the subsurface of oil and gas reservoirs. It is recognized that, because of the nature of the geological and production data, and the limitations of the numerical models to properly represent the true physics of real reservoirs, it is not possible to resolve uniquely and deterministically the underlying reservoir description. Thus the uncertainty in the resolution of the subsurface model translates into the uncertainty in the flow predictions (forecasts), which are one of the critical inputs to the reservoir management decision making processes. The importance of constraining the reservoir models to the observed production data (history match) is that it reduces the uncertainty in the reservoir model description and consequently it reduces the uncertainty in the forecasts. Unfortunately incorporating production data information into the reservoir model is not a simple task. It increases the complexity and difficulty of reservoir property estimation over the case of using geological data alone. The history match problem in reservoir simulation is not new; researchers and practitioners have been working in this area for at least 40 years and this has resulted in an extensive literature. Most of the published work relates to the use of automatic/assisted history match algorithms with the goal of finding efficiently a single good solution to the history match problem. History matching methods based on implementations of gradient based search algorithms proved to be very successful. However, history match methods with the goal of identifying multiple solutions, which are actually what is needed to make probabilistic forecasts, were not extensively explored, most likely because of the assumption that the computational cost would make them impractical to deal with the real field problems. Recent work in academics and industry is beginning to address the practical issues of dealing with the non-uniqueness in real field history match problems. The development of rapid, efficient and accurate computational methods and of associated computer IPTC 10751 History Match and Associated Forecast Uncertainty Analysis—Practical Approaches Using Cluster Computing J.L. Landa, SPE, Chevron Energy Technology Co., and R.K. Kalia, A. Nakano, K. Nomura, and P. Vashishta, U. of Southern California