Abstract: The prediction accuracy of the traditional coal spontaneous combustion temperature prediction model is poor. The requirement of parameter selection for the prediction model based on support vector machine (SVM) is high. And neural network-based prediction model is prone to overfitting. In order to solve the above problems, a prediction model of coal spontaneous combustion temperature based on random forest algorithm is proposed. The model uses the coal spontaneous combustion temperature program experiment to select O2 concentration, CO concentration, C2H4 concentration, CO/ΔO2 ratio and C2H4/C2H6 ratio as coal spontaneous combustion warning index data, processes the index data and divides the data into learning set and test set. The learning set is sampled to form a decision tree and split according to the optimal characteristics of the decision tree to form a random forest. The parameters of the random forest algorithm are optimized by the mean square error value and the determination coefficient (R2) to construct the random forest model. The test set data is input into the trained random forest model to obtain the prediction result of coal spontaneous combustion temperature. The model comparison results show that compared with the coal spontaneous combustion temperature prediction model based on the particle swarm optimization-back propagation(PSO-BP) neural network algorithm and the support vector machine algorithm, the R2 value in the random forest test phase is 0.869 7, the R2 value in the PSO-BP test phase is 0.783 6, and the R2 value in the SVM test phase is 0.835 0. The results shows that the prediction model of coal spontaneous combustion temperature based on RF algorithm can predict coal spontaneous combustion temperature more accurately and has strong robustness and universality. The model solves the problem that the prediction model of coal spontaneous combustion temperature based on PSO-BP neural network algorithm and the prediction model of coal spontaneous combustion temperature based on SVM algorithm are prone to overfitting.