The accurate prediction of coal temperature plays a vital role in preventing and controlling the spontaneous combustion of coal in coal mines. In this study, a large-scale 2-ton experimental furnace was constructed to implement an oxidation experiment by using a Dafosi coal sample. The aim was to simulate the spontaneous combustion of coal at low temperatures (< 100 degrees C). A random forest (RF) approach based on the oxidation experiment was proposed to predict coal spontaneous combustion, which exhibited satisfactory results. Moreover, to verify the performance and effectiveness of the RF approach in a practical application, a long-term in-situ observation test was conducted in a fully mechanized caving mining face (the Dafosi Coal Mine), where the in-situ data were employed to establish the RF model. Methods such as back-propagation neural network (BPNN) and multiple linear regression (MLR) were also adopted and compared with the RF model. The results indicated that the MLR model had the least reliable predicted results, regardless of whether the model was based on the oxidation experiment data or the in-situ data. This demonstrated that linear regression methods are not ideal for determining the complicated relationship between the temperature and gaseous products of coal spontaneous combustion. The BPNN model exhibited the most reliable prediction results during the training stage; however, overfitting occurred in the training stage, and the predictive performance at the testing stage was poorer than that of the RF model. The RF model accurately predicted the temperature of coal spontaneous combustion when it was applied to the in-situ data, and almost no deviation existed in the predictive performance indicators between the training and testing stages. The modeling and application results suggested that the RF model, which possesses high precision, strong generalization ability, and sound practical performance is more suitable for the prediction of coal spontaneous combustion. This method can potentially be further applied as reliable approach for the assessment of intricate relationships through fuel and energy investigations.