Reservoir evaluation method based on explainable machine learning with small samples

Abstract

Reservoir classification and evaluation of fractured gas reservoirs are essential for optimizing development strategies and enhancing oil and gas recovery rates. In this study, we utilized geological and engineering parameters to construct new feature dimensions and applied the K-means clustering algorithm to classify reservoirs into three categories based on unobstructed flow rates. We developed a novel machine learning framework that integrates Explainable Artificial Intelligence (XAI), Synthetic Minority Over-sampling Technique (SMOTE), and Stacking models, addressing class imbalance in small sample datasets. This framework achieved a classification accuracy of 92 %, demonstrating significant improvements over traditional methods. Through global and local interpretability analysis using SHAP values, we identified the critical features influencing the model's predictions, enhancing transparency and practicality. Using data from the Bozi-Dabei Block in the Tarim Basin, we validated the accuracy and applicability of our approach. This framework not only deepens the understanding of complex reservoir characteristics but also optimizes reservoir classification accuracy, providing robust technical support for the efficient development of unconventional oil and gas resources.