Reservoir models selection based on volumetric and connectivity metrics

Geostatistical sampling and stochastic inversion algorithms allow generating a large number of reservoir models conditioned by the available borehole and geophysical measurements. The spatial variability of these realizations represents the uncertainty in the reservoir model. However, in many practical studies, it is unfeasible to run fluid flow simulations and predict the reservoir production for hundredths of reservoir realizations due to the computational cost of dynamic fluid flow simulations. For this reason, we propose a statistical method for the selection of a subset of reservoir models that still captures the variability of the full set but significantly reduces the number of model realizations used in a numerical fluid flow simulator. We propose two statistical approaches: the quantile-based method that uses the quantiles of the distribution of a predefined metric and the low-dimension clustering method based on the centroids of cluster analysis in a low dimensional space generated using the multidimensional scaling based on the model dissimilarities. We illustrate both approaches using two metrics a volumetric metric that quantifies the total amount of hydrocarbon in place and a connectivity metric that defines the connectivity path between boreholes in the reservoir. The final result is the subset of n<N realizations (where N is the initial number of realizations), that are representative of the reservoir model uncertainty and variability. We test and validate the proposed approach on an idealized synthetic reservoir model representing a two-dimensional clastic reservoir with a high porosity channel saturated by oil within a non-permeable shale background, with two wells, an injector and a producer. We assume that 500 models are available from geostatistical simulations, and we select 5 models that represent the reservoir uncertainty. For illustration purposes, this example does not include seismic data conditioning; however, the proposed methodology can be extended to any set of models generated from geostatistical simulations, stochastic inversion of seismic data, or process-based modeling.