Abstract: Accurate water level prediction can provide scientific basis for flood control and disaster reduction and water resource management in lake area. In this paper, the IPSO improved particle swarm optimization algorithm and EGA elite genetic algorithm were introduced to optimize the LSTM long short term memory neural network structure. The improved IPSO-EGA-LSTM model was used to predict the water level of the gauging stations in Dongting Lake area under the 1d forecast period. The accuracy of the model was compared with the LSTM, GRU and BP neural network models. The prediction accuracy of the model was evaluated under the longer forecast period (3d, 5d and 7d). Three kinds of model input conditions were further set, corresponding water level prediction schemes were put forward, and the prediction accuracy of each forecasting scheme (direct forecasting, synchronous forecasting, rolling forecasting) under different forecast periods was explored. The results showed that the IPSO-EGA-LSTM model was better than the traditional neural network model in predicting the water level of Dongting Lake. It effectively captured the variation characteristics of water level of Dongting Lake in different forecast periods. The Nash efficiency coefficient (NSE) in 1d forecast period was greater than 0.998, and the NSE was still greater than 0.9 in long forecast period. The three prediction schemes under different input conditions demonstrated better prediction effects, among which the synchronous prediction scheme had better performance than the direct prediction and rolling prediction under the long prediction periods.