Dual-warning model for coal spontaneous combustion temperature prediction and risk classification based on BO-LightGBM

This study develops a dual-warning model for coal spontaneous combustion temperature prediction and risk classification based on BO-LightGBM. Five gas features (O₂, CO, C₂H₄, CO/ΔO₂, and C₂H₄/C₂H₆) were selected via Pearson correlation analysis, and six temperature-based risk levels were defined to characterize different combustion stages. The BO-LightGBM model was constructed using five-fold cross-validation and early stopping, with hyperparameters optimized separately for regression and classification tasks. Compared with grid search and random search, Bayesian Optimization significantly improved efficiency—reducing parameter optimization time by 91.2 % and 60.0 % in regression, and by 97.9 % and 85.1 % in classification, respectively. Performance comparisons further validate the superiority of BO-LightGBM. The model yielded a MAE of 10.9844, a MAPE of 7.57, a RMSE of 22.5728, and an R² of 0.9129 in regression tasks, and achieved an accuracy of 90.00 percent, a recall of 0.9000, a precision of 0.8920, and an F1 score of 0.8959 in classification tasks. These results confirm that BO-LightGBM offers notable advantages in both predictive accuracy and computational efficiency, providing a reliable early warning tool for spontaneous combustion risk assessment in coal mines.