Dual-Objective Prediction and SHAP Interpretability of Coal Spontaneous Combustion Based on Feature Attention and Heterogeneous MMoE

Coal spontaneous combustion involves highly nonlinear coupling among gas indicators, posing challenges for feature extraction, multi-objective prediction, and model interpretability under complex environmental disturbances. To overcome these limitations, we propose a feature-attention heterogeneous multi-gate mixture-of-experts (FA-HMMoE) model – a dual-objective prediction model for synchronously forecasting coal temperature and CO concentration. Drawing on temperature-programmed experimental data from the Dongtan coal mine, Spearman’s rank correlation analysis was applied to isolate four critical gas indicators tightly linked to temperature variations and CO emissions: O2, C2H4, the CO/ΔO2 ratio, and the C2H4/C2H6 ratio. An input-end feature attention module enhances key precursor signals, while a heterogeneous base composed of parallel fully connected and cross networks captures the high-order nonlinear interactions among these multidimensional gas indicators. We also incorporate an asymmetric joint loss function during optimization to prevent gradient dominance and negative transfer driven by dimensional differences. SHapley Additive exPlanations (SHAP) analysis is used to quantify feature contributions and relate model decisions to physicochemical reaction mechanisms. Five-fold cross-validation confirms the model’s robustness: the coefficients of determination (R2) for predicting coal temperature and CO concentration achieved 0.9393 ± 0.0209 and 0.9741 ± 0.0323, respectively, while the temperature RMSE was reduced to 17.8110 ± 5.0170. An additional leakage-control experiment excluding CO/ΔO2 retained R2 values of 0.9257 and 0.9558 for temperature and CO concentration prediction, respectively. The FA-HMMoE model achieved the highest R2 and lowest RMSE for both prediction targets among the compared models, although its CO prediction MAPE remained higher than that of several baselines.