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 long-term in-situ observation experiment was conducted in a fully mechanized caving face of the Dafosi Coal Mine, where the in-situ data of gases and temperature were obtained. Two machine learning approaches, random forest (RF) and support vector machine (SVM) were introduced and compared for predicting coal spontaneous combustion based on the in-situ monitoring data. The particle swarm optimization (PSO) was employed to optimize the RF and SVM by finding their optimal hyper-parameters. Principal component analysis (PCA) was used to transform the original input data into a new dataset of uncorrelated variables, reducing dimension for input variables. The results indicated that regardless of whether the models with or without PCA, the RF model was more robust than the SVM model and less affected by its own parameters, while the SVM model was highly sensitive to its parameters. Although the PSO could find the optimal hyper-parameters of the RF model, the RF model with default parameters could also accurately predict coal spontaneous combustion and possess satisfactory generalization. However, the predictive performance of the SVM model was dramatically improved in predicting after the PSO optimization. Moreover, the models with PCA also showed the above characteristics. These results suggest that both the RF and SVM methods can be used to predict coal spontaneous combustion, while the RF method can obtain accurate predictions without special parameter settings, it is more suitable for practical applications and can potentially be further employed as a reliable method for the determination of complicated relationships.