Abstract: To investigate the gas and temperature variation laws during coal spontaneous combustion oxidation and achieve accurate prediction of high-temperature points in coal spontaneous combustion, coal samples from Dafosi Coal Mine were collected for experimental research on coal spontaneous combustion. The characteristic indicators during coal oxidation were analyzed, including high-temperature point migration, gas volume fraction changes, oxygen consumption rate, and gas production rate. A random forest (RF) model for nonlinear prediction of coal temperature was established and validated using on-site monitoring data. The results showed that coal temperature rise exhibits staged variation characteristics and oxygen-deficient oxidation properties, with continuous temperature increases observed even when oxygen volume fraction remains between 4% and 5%, and rapid oxidation reactions still occurred under stable airflow conditions. Both oxygen consumption rate and gas production rates generally follow exponential growth trends, showing slow increases before reaching critical temperature and rapid acceleration after exceeding the dry cracking temperature. During coal oxidation heating, high-temperature points within the furnace demonstrate dynamic migration towards the air inlet, gradually shifting from the initial 45 cm position at the beginning of the experiment to the 5 cm position near the air inlet. The RF model demonstrated significantly superior predictive performance compared to support vector regression (SVR) and multiple linear regression (MLR) models under equivalent conditions, with prediction results showing mean absolute rercentage error (MAPE) <2.39% and coefficient of determination (R²) >0.99, while SVR and MLR models exhibited MAPE values exceeding 5%. Notably, the MLR model performed the worst, with a MAPE of 9.53% during the testing stage, highlighting the inadequacy of linear models in capturing the nonlinear relationships between gas products and temperature in coal spontaneous combustion. Further validation using on-site monitoring data confirmed the superior performance of RF model, achieving MAPE of merely 1.796% during testing stage, compared to 3.825% and 5.169% for SVR and MLR models, respectively.