<p>During coal spontaneous combustion (CSC), various gases are produced, which serve as quantitative indicators for assessing temperature distribution. A back propagation neural network (BPNN) model was developed to predict coal temperature based on gas concentration and spatial location, utilizing large-scale CSC experiments. The results indicated that O₂ concentration showed a phased decline over time, with notable changes in O₂, CO, and CO₂ levels beginning on the 42nd day. Variations in O₂ content within the coal led to uneven temperature distribution, and the high-temperature region gradually shifted toward the inlet side as spontaneous combustion progressed. Based on changes in maximum temperature and indicator gas concentrations, the coal's natural ignition process can be divided into three stages: slow oxidation, accelerated oxidation, and intense oxidation. The BP neural network model—incorporating gas concentrations at different time points and spatial distances of high-temperature regions—exhibited accurate predictive performance, with an overall <i>R</i>² of 0.9964. After optimization with the PSO algorithm, the RMSE was reduced to 2.57, further enhancing the model’s performance and generalization ability. This approach provides a valuable reference for predicting the progression of spontaneous coal combustion temperatures.</p> Graphical abstract <p></p>