Migration characteristics and prediction of high temperature points in coal spontaneous combustion

The determination and quantitative prediction of the high temperature points of coal spontaneous combustion in the gob is crucial for preventing coal spontaneous combustion fires. In this study, a large scale 2-ton experimental furnace was used to investigate the spontaneous oxidation process of coal and to explore the migration characteristics of the high temperature points of coal spontaneous combustion within the experimental furnace. Furthermore, a Bayesian-optimized K-nearest neighbors (BO-KNN) model for the quantitative prediction of coal spontaneous combustion temperature was established, and the model was compared with support vector regression (SVR) and partial least squares (PLS) models. The accuracy and robustness of the BO-KNN model were verified using in-situ monitoring data. The results show that the mean absolute percentage errors (MAPE) of the KNN, SVR, and PLS models during the testing stage were 2.065%, 6.893%, and 7.909%, respectively, and they reduced to 1.355%, 2.656%, and 7.503% after Bayesian optimization, respectively, indicating that the KNN model outperforms both the SVR and PLS models in terms of prediction accuracy. Moreover, the BO algorithm significantly enhances the predictive performance of the SVR model, while the improvement in the PLS model is minimal. The overall prediction performance of the PLS model is the lowest, suggesting its difficulty in capturing the nonlinear relationship between the gas products and coal temperature of coal spontaneous combustion. To further validate the practical performance of the KNN method, a long-term in-situ observation was conducted in the gob of the fully mechanized caving face at the Dafosi Coal Mine in Binzhou City, Shaanxi Province, China. Based on the in-situ data, KNN and BO-KNN models were developed, which demonstrated MAE < 1.033 °C, MAPE < 3.190%, and R2 > 0.912 during the training stage, and MAE < 0.902 °C, MAPE < 2.826%, R2 > 0.916 during the testing stage. These results confirm the high accuracy, strong generalization ability, and robust reliability of the KNN algorithm, making it highly suitable for the precise prediction of coal spontaneous combustion temperatures.