An interpretable machine learning model for optimization of prediction index gases in coal spontaneous combustion

Abstract

Early warnings of coal spontaneous combustion (CSC) have become urgent problems for coal enterprises. Existing approaches are designed to enhance the accuracy of CSC prediction. Improving the interpretability of the model is another important issue besides improving the prediction accuracy. Therefore, an interpretable machine learning framework based on RF (Random Forest) and SHAP (SHapley Additive exPlanations) is proposed to optimize prediction index gases. The data obtained from temperature-programmed experiments using coal samples from #5, #7, #8, #9, and #12 coal seams in Fangezhuang Mine are implemented to verify the proposed framework. CO, O₂/CO, CO/CO₂, CO/O₂, $CO/\Delta O_2$, $\Delta O_2$, $\Delta O_2/\Delta C{O}_2$, $C_2{H}_6/C{O}_2$, C₂H₄, CO₂/O₂ are selected, which is explained the rationality of the selected indicators using SHAP, practical experience, and related theories. Comparison of results using different machine learning models and different parameter optimization approaches showed the accuracy of the model affects the interpretation of the results. Finally, through the ablation experiment, the R² of RF, XGBoost, and Linear Regression model before feature removal was 0.98, 0.95 and 0.9, the model accuracy decreased significantly after the deletion, which showed the optimal prediction performance of RF, and the importance and validity of the selected indicators were verified using SHAP interpretation.