Abstract: The prediction model of residual coal spontaneous combustion temperature in goaf is weak of extrapolation and generalization at present. The main reason is that the number of model training data sets is small and the data distribution characteristics are not obvious due to the monitoring process and methods. In this paper, the WGAN-GP model is used to generate virtual samples and reconstruct and enhance the measured data sets. Through parameter correlation calculation, the variation of correlation coefficient among parameters of the generated virtual data set is not more than 15%. The results show that the prediction performance of each model is improved, among which the R2 index GA-BPNN model is improved by 12%, GA-SVM model is improved by 4%, RF model is improved by 3%. MAE index was decreased by 0.67 ℃ for GA-BPNN model, 0.54 ℃ for GA-SVM model and 0.33 ℃ for RF model. The RMSE index was decreased by 0.41 ℃ in GA-BPNN model, 0.46 ℃ in GA-SVM model and 0.39 ℃ in RF model. The enhanced and expanded data sets can improve the performance of the three prediction models, among which GA-BPNN model has the greatest improvement.