Abstract: Water temperature is one of main influencing factors on the water ecosystem of reservoirs. It is very necessary to predict the future water temperature for the uniform distribution in the reservoir area. Particularly, the deep learning is ever increasing in reservoir ecology. In this study, an improved long short-term memory (LSTM) model was proposed to simulate the water temperature in reservoir. A correlation analysis module was embedded to automatically filter the feature inputs of the model. In addition, the output dimensions module was also optimized. As such, the high accuracy was achieved to realize the rapid simulation and real-time prediction of the water temperature distribution in the reservoir area. Some suggestions were also provided for the reservoir scheduling and operation. The improved LSTM model was then applied to verify the simulation in the Xiluodu Reservoir project. The results show that the maximum root mean square error of the improved LSTM model was 0.63, and the minimum Nash-Sutcliffe efficiency coefficient was 0.96, indicating the better performance. The improved model was accurately captured the long-term dependence relationship in the data. The statistical parameters were included the average prediction errors and relative error, as well as the maximum and minimum water temperatures for the typical layers of the reservoir in all seasons. It was found that the maximum prediction error was 0.67 ℃, which was 0.10 ℃ for the minimum. The maximum and minimum relative errors were 4.17% and 0.45%, respectively. Among them, the maximum relative error was significantly less than 10%, fully meeting the expected level of the model. Meanwhile, both the maximum prediction error and the maximum relative error of water temperature occurred in the highest water temperature in spring (March-May). While the relative errors were less than 3% in the prediction of water temperature for the typical layer in the rest seasons. A comparison was also made on the improved LSTM and environmental fluid dynamics code (EFDC) model. There was the better consistence in the distribution of water temperature in the reservoir over time. The inter-annual error was -1.19-1.04 ℃ for the surface water temperature, with the relative errors of -5.6%-4.9%. In terms of the middle water temperature, the inter-annual error was -1.06-1.68 ℃ with the relative errors of -8.3%-6.2%. The inter-annual errors in the bottom water temperatures were ranged from -1.28-1.07 ℃, with the relative errors of -5.7%-7.1%. All relative errors were less than 10%, indicating the better prediction of the improved model. The computational efficiency was improved greatly. The simulation time was also shortened to a few hundred seconds, compared with more than a few days in the EFDC model. The rapid, real-time and accurate prediction of the water temperature distribution was realized by improved LSTM model. However, the number of output layers was limited by the measured data, leading to the large fluctuations of water level during prediction. Some suggestions were also offered on the number and location of output layers, segmental prediction, and the attenuation mechanism of the water temperature transfer.