Abstract: The surface subsidence caused by coal mining poses a severe threat to the ecological environment and personnel life safety, therefore, accurate prediction of surface subsidence is crucial for decreasing risks. At present, the adopted probability integration method has problems of low accuracy and difficulty in determining parameters in calculations. Although machine learning algorithm is introduced for improvement, this method is still prone to getting stuck in local optima and has a relatively slow convergence speed, which affects the accuracy of prediction. For this reason, a probability integration method for settlement prediction method based on GA-XGBoost is proposed, the optimal values of learning rate, maximum tree depth, and minimum leaf node weight of XGBoost model are obtained through GA algorithm; The excellent nonlinear fitting ability of XGBoost model is utilized to improve its predictive performance; Compared and analyzed with GA-BP neural network model and XGBoost model, GA-XGBoost model shows better determination coefficient R² (0.95) and root mean square error (0.008) than GA-BP neural network model and XGBoost model, with the smallest error in prediction, and GA-XGBoost model is applied to the 15210 working face of Youzhong Coal Industry. The results show that the error of GA-XGBoost model is smaller than that of GA-BP neural network and XGBoost model, and it achieves excellent effects in engineering practice.