In order to effectively predict the continuous water level in a certain period in advance, a short-term water level prediction model based on sequence to sequence (Seq2Seq) was proposed. The model used a long short-term memory (LSTM)network to encoder the historical water level sequence into a context vector, and another LSTM network was used to decode the context vector to predict the target water level sequence. Taking the Liuxi River as the research object, the water level prediction model was established for different prediction lengths, and compared with LSTM model and artificial neural network (ANN) model. The results show that the highest Nash-Sutcliffe efficiency coefficient predicted by the Seq2Seq model for continuous 6 h, 12 h and 24 h is 0. 93, 0. 90 and 0. 85, respectively. When the prediction length is 6h, the performance of the LSTM and Seq2Seq model are similar, and the ANN model has lowest precision. When the prediction length is 12h and 24h, the Seq2Seq model has better performance and faster convergence speed compared with the LSTM model and ANN model.