Geological and dynamic similarity for reservoir state prediction by well connectivity

Abstract

For decades, computational reservoir modeling remained the main instrument for assessing reservoir state during the operations, and for the forecasting of future production rates. The recent development of reduced-order physics modeling and pure data-driven solutions made it possible to monitor and forecast the reservoir state in a more prompt way. In this work, we propose a new data-driven approach aimed at assessing inter-well connectivity. The solution is based on a machine learning (ML) forecasting model for production parameters (production rates, bottom-hole pressure) and the extraction of feature importances by additive explanation. Resulted inter-well connectivity coefficients are used as edges weights in well connectivity graphs. The validation of the proposed solution was performed by the comparison with Capacity Resistance Model performance and static geological similarity between wells derived from well logs analysis. Two case studies on benchmark reservoir models from clastic and carbonate environments are presented to ensure reliability of the proposed solution.