The prediction of water level for important stations of large reservoirs is an essential issue in flood control. At present, hydrodynamics method is mainly used to calculate the water level. However, due to the high accuracy requirements of input boundary conditions, which is difficult to meet under real-time dispatching, the calculation error is likely to be large. In view of this, we present a recurrent neural network model to mine the knowledge from historical data of reservoir operation and to learn the mapping relationship between inlet flow (including main stream and tributaries) and the water level from in front of the dam to the target station. This method does not use terrain data to avoid the effect of systematic error, thereby reducing the requirement for input boundary accuracy on one hand and improve prediction accuracy on the other. In the experimental part, we compute the water level of Changshou Station of Three Gorges Reservoir. The data sets (including training, validation and test sets) consist of the operation data from 2009 to 2019. When the water level is close to the land requisition line, the prediction error is within±0.4 m, which is better than the result of water dynamic model. Therefore, the proposed method can meet the accuracy requirement of the real-time scheduling of water level.