Characterizing reservoir connectivity and forecasting waterflood performance using data-driven and reduced-physics models

Closed-loop reservoir management requires representative models that can be updated quickly. Such kind of models that can be used for characterization of reservoir connectivity would greatly contribute to decisions in waterflooding operations. In heterogeneous reservoirs that are characterized with thief-zones/fractures or flow-barriers, that may enhance or prevent connectivity, it becomes challenging to construct geological models with both reasonable accuracy and computational efficiency. In this study, practical and computationally-efficient methods are presented to first accurately characterize the connectivity between producer and injector wells, and then to forecast the waterflood performance. Two classes of modeling approaches are evaluated in terms of their applicability to characterize the connectivity: 1) Data-driven modeling - artificial neural networks (ANN), in which there is not any presumed functional relationship and connectivity is derived from the weights on connection links of neural networks; 2) Reduced-physics modeling - capacitance-resistance models (CRM), in which physics are incorporated with certain assumptions and connectivity is derived from the fractional flow parameter in the formulation. In both approaches; a producer-based control volume is considered to focus on producers and their connectivity to different injectors, and both models are trained with injection/production histories. Connectivity values derived from these methods are fed into a short-term, pattern-based forecasting methodology, which quantified ultimate recoveries at the desired oil-cut values. Two synthetic reservoir models are used to test the applicability of aforementioned methods; which are characterized with high-permeability streaks and flow-barriers. These key features that affect connectivity are identified by both modeling approaches with reasonable accuracies, but with varying interpretations of fluid movement in the reservoir. Pattern-based forecasting methodology resulted in ultimate recovery predictions very close to the numerical-model based results, with errors within 3%. The workflow presented is quick and reliable to be implemented in closed-loop reservoir management protocols, which requires models that can be updated easily with recent data, and can be used in decision making processes with confidence.