Research on Temperature Distribution Prediction in Coal Mine Goafs Based on Multi-Field Coupling and Machine Learning Framework

Accurate and rapid prediction of coal spontaneous combustion temperature is crucial for preventing mine fires, reducing casualties, and ensuring property safety. Machine learning based numerical prediction techniques are increasingly esteemed by practitioners. However, existing machine learning based models struggle to efficiently predict multi-point temperature distributions under variable working conditions. This study established a multi-field coupled model based on Opt-XGBoost technique to predict temperature field distributions in goaf areas under different mining parameters. By extracting gas component distributions and mining status parameters as model inputs, we constructed nonlinear mapping relationships between these variables and spatial temperature patterns. A novel multi-objective prediction method integrating cluster analysis, parameter optimization, and ensemble learning was developed. Using k-means clustering, temperature monitoring points were dimensionally reduced from 16,765 to 427 representative locations. Multi-output algorithms synchronously predicted temperatures at these positions, with full-field distribution maps reconstructed from discrete predictions. The optimized model achieved an average coefficient of determination (R2) of 0.9762 and Mean Squared Error (MSE) of 0.462, significantly outperforming baseline results (R2 : 0.9269). High-temperature zones in prediction maps closely matched simulated temperature fields, with complete distribution predictions completed within two seconds, meeting real-time monitoring requirements. These results demonstrate the approach’s dual advantages in prediction accuracy and computational efficiency, providing a reliable predictive tool for spontaneous coal combustion risks in mining operations.