Abstract: Water temperature is one of main factors affecting the water ecosystem of reservoirs. It is important to understand the distribution of water temperature in the reservoir area for protecting the reservoir ecosystem. Predicting the future water temperature changes is also of great value. With the increasing application of deep learning algorithms in reservoir ecology, this study proposed an improved long short-term memory (LSTM) model, which embedding a correlation analysis module to filter the feature inputs of the model automatically. In addition, the output dimensions module was also optimized. Through that, the aim of rapid simulation and real-time prediction of the water temperature distribution in the reservoir area was achieved, which under the premise of the prediction accuracy guaranteeing. The study would provide guidance suggestions for reservoir scheduling and operation. The improved LSTM model was applied to the Xiluodu Reservoir project, the results show that the maximum value of the root mean square error of the improved LSTM model is 0.63, and the minimum value of the Nash-Sutcliffe efficiency coefficient is 0.96, which indicates that the model had a better performance overall. It also shows that the improved model can accurately capture the long-term dependence relationship in the data. By the statistical results of the prediction error values and relative error of the average, maximum and minimum water temperatures for the typical layers of the reservoir in all seasons, it is found that the maximum value of the prediction error value is 0.67 °C, which is 0.10 °C for the minimum value. It is also calculated that the maximum and minimum relative errors are 4.17% and 0.45%, respectively. That is, the maximum relative error is significantly less than 10%, which reached the expected level of the model. By further analysis, the statistical results meanwhile reveal that both the maximum prediction error value and the maximum relative error of water temperature occurred in the highest water temperature results in spring (March-May), while the relative errors of the prediction of water temperature for the typical layer in the other seasons were less than 3%. By comparing the results of improved LSTM model and environmental fluid dynamics code (EFDC) model, it could be found that the water temperature distribution and value changes rule in the reservoir of two models were basically consistent over time. The inter-annual error value is -1.19～1.04 °C 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 value is -1.06～1.68 °C and the relative errors is -8.3%～6.2%. The inter-annual errors in bottom water temperatures ranged from -1.28～1.07 °C, with relative errors of -5.7%～7.1%. The above results show that all relative errors are less than 10%, which indicated that the predicted results could be accepted. Compared with the simulation time of the EFDC model which was more than a few days, the improved model shortened the calculation time to a few hundred seconds, and the computational efficiency was greatly improved. In other words, the fast and real-time accurate prediction of the water temperature distribution is realized by improved LSTM model. However, there are also some limitations and deficiencies for the LSTM model. For example, the number of output layers will be limited by the measured data, and it is difficult to predict the water temperature distribution in reservoirs with large water level fluctuations. To solve the above problems, the present study puts forward some optimization suggestions on the reasonable selection of the number and location of output layers, segmental prediction, and the attenuation mechanism of the water temperature transfer.