Prediction of Shale Gas Well Productivity and Analysis of Key Factors Based on Interpretable Machine Learning

Summary Accurately predicting shale gas production dynamics is crucial for optimizing development plans, reducing costs, and advancing technological progress. Traditional prediction models have problems such as high input dimensions, high computational complexity, easy neglect of key features, and poor interpretability. On this basis, we propose an interpretable machine learning (ML) framework for shale gas production to improve model performance and quantitatively evaluate the contribution of dominant factors. Initially, six ML models were developed, and after data analysis, eight features were used as inputs, and hyperparameters were adjusted through Bayesian optimization. Subsequently, the recursive feature elimination cross-validation (RFECV) algorithm is used to optimize key features and reduce model dimensionality. Finally, game theory was combined with three optimal ML models, and the Shapley (SHAP) method was used to explore the global and local effects of various factors and conduct model stability analysis. The complete workflow was validated using actual production data from a shale gas well in China. The results show that the performance of all models has been improved. The coefficient of determination (R2) of the light gradient boosting machine (LGBM) model increased by 11% (from 0.75 to 0.86), while the extremely randomized tree (ETR) model had the highest accuracy (R2 = 0.95). The SHAP analysis results confirmed the effectiveness of the RFECV method and revealed for the first time the interaction between features, resulting in highly similar conclusions and improved interpretability. The results of this study will help on-site experts to conduct more accurate and objective analysis and optimization of shale gas production with minimal data science expertise.