Improved Methodologies for Stochastically Forecasting Oil Recovery Processes

by
Francisco Jose Guevara Baro, Ph.D.

University of Texas at Austin, 1997
Supervisors: Mark A. Miller and Kamy Sepehrnoori

Accurate forecasting of the behavior of a reservoir recovery process requires knowledge of geological information, fluid properties, and petrophysical data. Because of the complexity of geological deposition and diagenesis, the spatial distribution of these properties is very complex. Moreover, this information is only known at well locations, which represent only a very small fraction of the entire reservoir. In addition, many other parameters (such as initial water saturation, residual oil saturation, relative permeability curve endpoints), are not strictly deterministically known. A single reservoir description to obtain a forecast of a reservoir process is thus not complete. Instead, a statistical forecast is more appropriate. This statistical treatment would imply that many reservoir descriptions should be generated. These reservoir images must usually be processed through a finite-difference simulator for predictive purposes, requiring unreasonably large amounts of computer time and human effort.

Two new methodologies for transferring uncertainty are presented in this study. The first methodology, called the combination algorithm, uses simple models (e.g., streamline, predictive or streamline finite-difference models) to rank many equiprobable sets of single-valued reservoir stochastic parameters and random permeability fields through an economic indicator value and a response parameter, respectively. Then only a few sets of these parameters and fields, corresponding to some cumulative probability values or selected points, are used to run the complex model, from which an approximate NPV cumulative distribution function is obtained.

The second methodology, called the ranking function algorithm, requires sensitivity analysis from the complex model to generate a function that plays the role of the simple model in the above methodology. Polymer flooding was chosen as the process to be studied. This process has sufficient complexity to be of interest, yet it is simple enough to provide early insight.

These two methodologies are tested at three different levels, at test level using a streamline model and predictive model, at full-process simulator level using a predictive model and a commercial simulator, and at field scale with the same models. The approximate cumulative distribution functions obtained at the three levels and for both methodologies were reasonably accurate to ones obtained from a direct Monte Carlo approach. The ranking function algorithm produced lower errors than the combination algorithm at only slightly higher cost, especially when sensitivity studies are to be conducted. Also, it is concluded that multiple algorithm realizations are required to achieve an accurate cumulative distribution function.

Back to theses index