Pore Structure Prediction from Well Logs in Deep Tight Sandstone Reservoirs Using Machine Learning Methods

In this study, deep tight sandstone was selected as an example to propose a complete method for predicting reservoir pore structure by capillary pressure curves and conventional well log data. This method pioneers the integration of grey relational analysis, principal component analysis, ensemble clustering, and deep neural networks to establish a systematic predictive framework for transitioning from conventional logging data to pore structure types. A total of 186 core data from three wells were used in this study. First, sensitive pore structure parameters from mercury injection capillary pressure data were extracted using grey correlation analysis and principal component analysis. Then, unsupervised clustering analysis was applied to classify the reservoir pore structures in the study area, dividing it into three categories. These category labels were combined with conventional well logs to create learning samples for a deep neural network (DNN) model developed to predict reservoir pore structure categories. The accuracy of the training set of the model reached 88.2%, while the accuracy of the testing set was 80.43%. Finally, the method was applied to field well log data. The results showed significant differences in pore structure classifications among gas layers, water–gas layers, and dry layers. This method is versatile, with its core components transferable to other deep sandstone reservoir studies, and can accurately predict the pore structure of tight sandstone reservoirs, which is critical for advancing the characterization of deep and complex oil and gas reservoirs.