Abstract: In order to reduce the damage caused by frequent irregular action to the hydropower stations of high water head, small storage capacity and poor regulation performance in the long-term operation process, maximize the use of the water head advantage to increase power production, and improve the efficiency and safety of hydropower station operation, a water level prediction method driven by mechanism and data is proposed.In this method, Back Propagation(BP) neural network and water balance mechanism are coupled with Particle Swarm Optimization(PSO) algorithm, in which the data-driven model provides the reference value and the water balance mechanism model corrects the rationality of water level trend.This method is applied to the water level prediction of Shaping Ⅱ Hydropower Station, and the prediction results of water balance prediction model, BP neural network prediction model and coupled model are compared and analyzed.The results show that the proposed coupled model effectively avoids the accumulation error of the mechanism model and the un-constancy of the data-drive model.Compared with the water balance prediction model and the BP neural network prediction model, the coupled model has higher prediction accuracy and practicability.The mean absolute percentage error and goodness of fit are 0.001 3 and 0.97,respectively, and the prediction amplitude is closer to the real water level.The research results can provide theoretical basis for hydropower station to respond in advance to short-term water level changes.