Data Mining Methodologies to Reduce the Uncertainty of Reservoir Selection

Abstract

Geostatistical based earth modeling can create hundreds of reservoir property realizations. The challenge is to select a few optimal models from these realizations for further analysis. Realizations are often ranked according to various criteria based on pore volume, connected or drained volume. Traditional workflows then select candidates from ranked distributions representing quantiles (e.g., P10, P50, and P90) of the reservoir volumetrics, which assumes contiguous reservoir volumes. However, such methods lack any physical geometrical information. Geostatistical reservoir modeling could result in realizations that have unconnected pore volumes equaling pore volumes of other realizations with connected pores. This paper discusses two novel ways to choose simulation models from multiple realizations. Before application of data mining techniques, a three-dimensional (3D) realization is collapsed into a one-dimensional array, so that each element of the array maps to one single i, j, k indexed cell. This array carries the spatial property information. A matrix is then created from multiple realizations. Each column represents an individual simulation and each row represents a single 3D grid cell. The first method introduced is two-way clustering of both columns (R-mode) and rows (Q-mode). It is a new method that quantitatively evaluates cell by cell of a grid throughout the realizations. During cluster analysis, the spatial property location of the cell is carried in a distance calculation. R-mode clustering shows the similarity of different simulations. Domain experts can first select candidate simulations from different clusters and then compare the results. Conversely, Q-mode clustering shows the grouping of different cells. The location and connectivity of cells can be mapped in addition to evaluating their statistical ranking. The second method uses classification and regression trees to select the reservoir simulation that strongly associates with other petrophysical properties or seismic attributes. It can associate with both numerical and categorical properties. Both methods were applied to a West Texas Permian basin reservoir. Through above novel quantitative evaluations, geologists and reservoir engineers can select optimistic to pessimistic realizations to aid in the economic assessment of the reservoir. They help reduce uncertainty when applying volumetric histogram cutoffs, particularly in shale plays where the reservoirs may not be contiguous.

AAPG Datapages/Search and Discovery Article #90259 ©2016 AAPG Annual Convention and Exhibition, Calgary, Alberta, Canada, June 19-22, 2016