Identification Method of Remaining Oil Potential Area Based on Deep Learning

Abstract

Efficiently classifying potential areas of remaining oil is essential for enhancing the recovery in high water cut reservoir. The distribution of remaining oil is complex and challenging to mobilize due to temporal evolution and spatial variation in long-term waterflood development. Currently, reservoir classification relies on manual experience and unsupervised machine learning, both of which have limitations. Manual classification is constrained by human understanding, leading to inaccuracies, while unsupervised learning lacks adherence to reservoir theory, resulting in a possible lack of physical interpretability. This article introduces the Beluga whale optimization and the improved temporal convolutional network (BWO-ITCN) model, a novel prediction classification model that combines intelligent classification with reservoir theory constraints. The structure of the ITCN model was improved by changing the serial structure to the parallel pooling structure, in order to extract the features of time series data. The BWO-ITCN model incorporates expert experience by considering five dynamic and one static indicators for evaluating potential areas. It aims to identify remaining oil potential areas by learning the evaluation indicators in spatial variations and temporal evolution. The BWO algorithm enhances the classification precision by optimizing hyperparameters, particularly, blending samples. The experimental results demonstrate that the BWO-ITCN model achieves an accuracy of 94.25%, a precision of 94.1%, a recall rate of 93.82%, and an F1-score of 93.83%. Notably, the overall accuracy of a spatiotemporal model is higher than a nonspatiotemporal model. This autonomous classification model effectively addresses the challenges in classifying potential areas, simplifies the process, and offers valuable insights for development.