Abstract: A high-precision temperature prediction model for coal spontaneous combustion under complex oxidation conditions is developed to enhance stage identification and support early warning in underground coal mines. Programmed heating and simultaneous thermal analysis experiments are performed under varying oxygen concentrations and heating rates. Characteristic temperature points (T_C1—T_C7, T₁—T₆) are extracted, and the oxidation process is divided into seven stages according to their evolution and functional group responses. The process is further integrated into four macro intervals for calculating apparent activation energy and enthalpy change using the Coats–Redfern method. These thermal and gas features are used as inputs for temperature prediction models based on XGBoost and GBR algorithms. SHAP analysis is applied to interpret feature contributions. Results show that T_C6 and T_C7 are shifted significantly toward higher temperatures as oxygen concentration decreases, with deviation rates of −2.0 ℃/% and −1.33 ℃/%, respectively. T₂ and T₄ are highly sensitive to increases in heating rate, with deviation rates of 2.85 ℃/( ℃·min⁻¹) and 2.83℃/(℃·min⁻¹). The thermal runaway threshold stage exhibits the highest activation energy (78.86 kJ/mol) and enthalpy change (74.16 kJ/mol), indicating intensified heat release. Superior predictive performance is achieved by the XGBoost model (R² = 0.999 6, MAE = 0.32 ℃), outperforming GBR. SHAP analysis confirms stage-dependent contributions of thermal and gas parameters, thereby improving physical interpretability. A reliable basis is provided for dynamic identification of combustion risk and for optimization of nitrogen injection and ventilation strategies in spontaneous combustion control.