Development of a deep learning-based model for predicting of dominant seepage channels in oil reservoirs

Abstract The cost of implementing improved oil recovery measures after the formation of dominant seepage channels is high, and the effectiveness is often not significant. By predicting the formation of dominant seepage channels and actively intervening in their early stages, it is possible to reduce development costs and improve oil recovery factor. Consequently, the prediction of dominant seepage channels has emerged as a key focus of research in reservoir engineering. This research introduces a novel method for predicting dominant seepage channels, termed Label Matrix of Seepage Channels Informer (LMSC-Informer), which integrates deep learning with reservoir engineering principles. It employs an evaluation method for the development of dominant seepage channels and a label matrix for seepage channels. Unlike existing approaches that primarily depend on geological and formation parameters, this method leverages reservoir production data, making it more accessible and versatile. An experimental prediction conducted in a reservoir in China demonstrated the practical effectiveness of the proposed methods, achieving a prediction accuracy of 73.9%.