Screening of a Probabilistic Database Using Stochastic Reasoning Driven by Surrogate Models

Waterflooding is among the oldest and perhaps the most economical of oil recovery processes to extend field life and increase ultimate oil recovery from naturally depleting reservoirs. During waterflood operations, water is injected into the reservoir to maintain a certain reservoir pressure as well as to push the oil in the reservoir towards the producing wells. Nowadays, any organization always has to strive for lean and efficient technologies and processes to maximize profit also when looking deeper into their reservoir portfolios in order to identify additional waterflooding opportunities. Time and information constraints can limit the depth and rigor of such a screening evaluation. Time is reflected by the effort of screening a vast number of reservoirs for the applicability of implementing a waterflood, whereas information is reflected by the availability and quality of data (consistency of measured and modeled data with the inherent rules of a petroleum system) with which to extract significant knowledge necessary to make good development decisions. A new approach to screening a large number of reservoirs uses a wide variety of input information and satisfies a number of constraints such as physical, financial, geopolitical, and human constraints. In a fully stochastic workflow that includes stochastic back-population of incomplete datasets, stochastic proxy models over time series, and stochastic ranking methods using Bayesian belief networks, more than 1,500 reservoirs were screened for additional recovery potential with waterflooding operations. The objective of the screening process is to reduce the number of reservoirs by one order of magnitude to about 100 potential candidates that are suitable for a more detailed evaluation. Numerical models were used to create response surfaces as surrogate reservoir models that capture the sensitivity and uncertainty of the influencing input parameters on the output. Reservoir uncertainties were combined with expert knowledge and environmental variables and were used as proxy model states in the formulation of objective functions. The input parameters were initiated and processed in a stochastic manner throughout the presented work. The output is represented by a ranking of potential waterflood candidates. The benefit of this approach is the inclusion of a wide range of influencing parameters while at the same time speeding up the screening process without jeopardizing the quality of the results.