Prediction of Temperature Intervals Driven by Multiple Indicator Gases during Spontaneous Coal Combustion and Oxidation

Coal spontaneous combustion (CSC) remains one of the major disaster-inducing factors in underground coal mines, and accurate temperature-interval prediction is essential for developing effective prevention and control strategies. In this study, temperature-rise experiments on CSC are conducted using coal samples from the No. 4 coal seam of the Songxinzhuang Coal Mine, operated by Sinopec Great Wall Energy & Chemical (Ningxia) Co. A self-developed system for screening and analyzing spontaneous combustion indicator gases is employed to monitor gas evolution during the oxidation process. A total of 200 data sets are collected from CSC heating experiments, through which the critical temperature (T 1) and spontaneous combustion temperature (T 2) are identified. The experimental results reveal that O2, CO, CH4, CO2, C2H6, C2H4, and their ratio combinations (e.g., CO/CH4, C2H6/CH4, C2H4/CH4, and C2H4/C2H6) exhibit pronounced temperature-responsive behavior throughout the heating process. Drawing on the stage-wise evolution patterns revealed by these indicators, we developed a multi-indicator, gas-driven temperature-interval prediction model. The model integrates the Gray Wolf Optimization (GWO) algorithm, Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM) network, and Quantile Regression (QR), with indicator gases as inputs and coal temperature as the output. The model performance is evaluated using both point prediction metrics (coefficient of determination (R 2), root-mean-square error (RMSE)) and interval prediction metrics (coverage width-based criterion (CWC), mean prediction interval coverage distance (MPICD)), and comparative analyses are performed against several deep learning benchmarks, including CNN-BiLSTM, BiLSTM, GWO-BiLSTM, multilayer perceptron (MLP), and GWO-MLP. The proposed GWO-CNN-BiLSTM-QR model achieves a point prediction R 2 of 0.9554 and an RMSE of 16.1455 on the test data set, significantly outperforming traditional models. For interval prediction, the model yields CWC and MPICD values of 0.2019 and 17.887 at the 95% confidence level and 0.1561 and 16.502 at the 80% confidence level, indicating tighter, more reliable prediction intervals and demonstrating superior accuracy and robustness. Furthermore, the model's generalization capability is validated using independent experimental data from the Dongtan Coal Mine. The results confirm its strong predictive performance, offering both theoretical insight and practical guidance for CSC forecasting, early warning, and risk mitigation. Overall, the proposed framework demonstrates promising potential for large-scale engineering applications in coal mine safety management.