In order to overcome the shortcomings such as insufficient information mining capability of conventional water level prediction methods and unclear mechanism of heuristic algorithms, an water level prediction method based on long short-term memory(LSTM)neural network is proposed. Direct monitoring data such as water level and unit output are used in this method and the middle errors caused by the indirect calculation of the outflow and inflow can be avoided, which improves the accuracy of water level prediction. A hybrid method based on the gradient descent algorithm and Broyden-Fletcher-Goldfarb-Shanno(BFGS)algorithm is used to train the model, and the step length is determined by the Wolfe-Powell line search method to accelerate convergence. The proposed method is used to predict the water level at the upstream and downstream of the Gezhouba Hydropower Station. The results show that this method can continuously predict the downstream water level for 6 hours and the upstream water level for 3 hours with high accuracy, providing technical support for the real-time scheduling of the Gezhouba Reservoir.