The change of river water level is affected by many complex factors.The water level data not only shows nonlinear characteristics,but also features with time sequence and complexity.Improving the accuracy of water level prediction is of great significance to river management,water conservancy construction,water resources scheduling,flood control and disaster reduction,and shipping safety.This paper makes use of the advantages of long-term memory neural network （LSTM）in dealing with long-time series problems,the advantages of support vector regression（SVR）in dealing with nonlinear data,and the advantages of particle swarm optimization（PSO）in adaptive global search.The pso-svr-lstm combination model is applied to the water level prediction of WanJiaBu section of Xiuhe river.The simulation results show that the prediction accuracy of the proposed pso-svr-lstm model is higher than that of LSTM model,SVR model and BP model.