Aiming at the problems of local optimality, poor interference resistance, and low prediction accuracy of the traditional seepage monitoring model for earth-rockfill dams, through optimization of the Extreme Gradient Boosting (XGBoost) algorithm (RUN-XGBoost algorithm) by the Runge Kutta optimizer (RUN) algorithm, a RUN-XGBoost model was constructed to obtain better seepage prediction results.The RUN algorithm is applied to improve the three main parameters of the XGBoost algorithm during the initialization of the population, which gives high validity to the prediction results.The overall convergence speed and prediction accuracy of the algorithm are improved by automatic searching for the optimal parameter.A stochastic variance factor is also introduced to enable the algorithm to exclude local minima and continue the search to obtain a globally optimal result.The validation results of engineering examples show that the RUN-XGBoost model has the advantages of simplicity, high efficiency, high prediction accuracy and robustness.