Water hysteresis is an important factor affecting the short-term optimal dispatch of cascade hydropower stations. How to construct a dynamic hysteresis model to accurately describe the upstream and downstream hydraulic connections of cascade hydropower has become an urgent problem to be solved in the short-term optimal dispatch of hydropower. This paper attempts to introduce the eXtreme Gradient Boosting (XGBoost) algorithm as a tool to express the dynamic water hysteresis between cascade hydropower stations. First, it uses the Shapley addictive explanation (SHAP) to evaluate the importance of each input feature and select the optimal feature set. Second, a dynamic hysteresis model based on an extreme gradient lifting tree is established. Grid search and cross-validation algorithms are used to select the model parameters. Finally, based on an actual case, comparative experiments with artificial neural networks, support vector regression and a fixed time-lag model are carried out. The results show that the constructed SHAP-XGBoost model can simulate the dynamic hysteresis relationship between hydropower stations more accurately, the average prediction accuracy of the proposed model is improved by at least 18% compared with other models, and it can best match the change trend of the measured sequence. It is also proved that the addition of input feature screening can further improve the model accuracy.