In order to further improve the accuracy of water level prediction, this paper proposes a Long Short Time Memory (LSTM) prediction model integrated with improved attention mechanism. The model divides the input sequence into the time sequence and the characteristic sequence. The attention mechanism is introduced before the LSTM network model to compute the attention of these two sequences separately. After intergrating, the LSTM network is able to adaptively select the most important input features according to their importance, and the parameters of the attention mechanism layer are obtained by a competitive random search algorithm.Thus, the robustness of the model is further enhanced. Finally, a prediction experiment is conducted on the water level data of Poyang Lake. The experimental results show that compared with the support vector regression (SVR) and LSTM model, the proposed LSTM model based on improved attention mechanism has better prediction accuracy and can provide technical support for water level prediction and precise regulation of water resources.