Abstract:
To more accurately capture the time-varying patterns of residual coal temperature in goaf and achieve precise prediction of coal spontaneous combustion temperature, this paper introduces an improved Whale Optimization Algorithm (WOA) incorporating adaptive weights and the Levy flight strategy, proposing the Adaptive Levy Whale Optimization Algorithm (ALWOA). The superior convergence speed and accuracy of ALWOA compared to WOA are validated using four benchmark functions. Based on this, ALWOA is utilized to optimize a Bidirectional Long Short-Term Memory network (BiLSTM), constructing a multi-parameter prediction model for residual coal oxidation temperature in goaf based on ALWOA-BiLSTM. Using eight indicators, including O?, CO, CO?, CH?, and CO/O?, from the 11301 working face of Hongqingliang Coal Mine as input features and temperature as the prediction target, the model was trained on 80% of the data and tested on the remaining 20%. The results show that the ALWOA-BiLSTM prediction model achieves a Mean Absolute Error (MAE) of 0.7208, a Root Mean Square Error (RMSE) of 0.9395, and a coefficient of determination (R2) of 0.8724, demonstrating high prediction accuracy. Compared to BiLSTM and WOA-BiLSTM models, the MAE is reduced by 25.24% and 29.00%, the RMSE by 13.20% and 20.74%, and the R2 is improved by 14.97% and 10.01%, respectively, validating the effectiveness of using ALWOA to optimize the BiLSTM model. Furthermore, the SHAP method was introduced for model interpretability analysis. Global feature importance ranking indicates that R1, R2, and CO are the most critical indicators affecting temperature prediction. Feature dependence analysis reveals significant threshold effects for both R1 and R2 (R1≈1.30, R2≈0.007), beyond which their driving effect on temperature intensifies sharply. Local interpretations visually demonstrate the contribution patterns of each indicator under different temperature states. The physical response mechanisms revealed by the SHAP analysis align closely with the mechanisms of coal spontaneous combustion oxidation, verifying the physical rationality of the model's learning and providing data-driven decision support for graded warning and precise prevention and control of coal spontaneous combustion in goaf areas.