The prediction of reservoir water level provides important decision support for its operation, flood control and water resources operation and management.Accurate and reliable prediction plays an important role in the optimal management of water resources. Hydrodynamic prediction method requires high input boundary conditions and large prediction error,which is difficult to meet the requirements of real-time water level regulation.Aiming at the nonlinearity, instability and complex temporal and spatial characteristics of reservoir water level data,a hybrid reservoir water level prediction model integrating adaptive Variational Modal Decomposition (VMD),Convolutional Neural Networks(CNN) and Gated Recurrent Unit (GRU) is proposed. Among them,VMD eliminates noise by decomposing the water level sequence,CNN is used to effectively extract the local features of water level data,and GRU is used to extract the deep time features of water level data.Taking the daily water level prediction of Shenwo reservoir as an example,compared with several related models,the analysis results show that the new model performs best in the selected evaluation indexes. Among them, the Root Mean Square Error is 0.1500, the Mean Absolute Error is 0.1063,the Mean Absolute Percentage Error is 0.1145,and the Nash–Sutcliffe efficiency coefficient is as high as 0.9953.In particular,the operation efficiency of GRU selected in this paper is significantly improved compared with Long and Short-term Memory network(LSTM). Therefore,the new model has high accuracy and high operation efficiency, and is more suitable for the real-time operation of reservoir water level.