Abstract: In order to obtain accurate and effective prediction and early warning of coal spontaneous combustion, a coal spontaneous combustion prediction (PSO-XGBoost) combined with particle swarm optimization algorithm and limit gradient boosting regression tree model (XGBoost) was proposed. Among them, the random sampling rate in the gradient boosting regression tree and the weight parameters of the minimum leaf node samples were optimized and combined by PSO algorithm. Based on 337 sets of data obtained from coal spontaneous combustion experiment in Dongtan Coal Mine, the volume fraction of O₂, the volume fraction of CO, the volume fraction of C₂H₄, the ratio between the volume fraction of CO and the volume fraction of remaining O₂, and the ratio between the volume fraction of C₂H₄ and the volume fraction of C₂H₆ were selected as indicators. Firstly, 70% of the total data was randomly selected to be divided into training set and 30% into test set. The inertia weight of PSO algorithm was designed to optimize the parameters of XGBoost and select the best combination of parameters to build the model. Then, the PSO-XGBoost model was compared with the prediction results of the standard XGBoost model, the random forest (RF) model and the gradient lifting tree (GBRT) model. The results show that the mean absolute percentage error (MAPE) of the training sample prediction results of RF, GBRT, XGBoost and PSO-XGBoost models are 2.256%, 2.423%, 0.276% and 0.072%, respectively, and the MAPE of the test sample prediction results are 7.246%, 8.816%, 7.594% and 6.860%, respectively. The results show that the accuracy of PSO-XGBoost model is better than that of RF model, GBRT model and XGBoost model. The PSO-XGBoost model is more suitable for prediction and early warning of coal spontaneous combustion.