Abstract: 　　River water temperature plays a central role in river ecological health assessment and water resources management. However, due to the limitation of observation conditions and equipment maintenance in historical monitoring, the water temperature data are generally missing and discontinuous, which limits the research on watershed thermal process and ecological response. We constructed the water temperature time series reconstruction machine learning model based on LSTM algorithm by taking the Zumuzu hydrological station in the upper reaches of Dadu River as the research object, and used the SVR model as the control. The time series samples were constructed by sliding window technology, and the time series of ten-day average water temperature from 2007 to 2020 were reconstructed by combining multi-source input factors such as temperature, flow and solar radiation. The results show that the average relative error ( MRE ) of the LSTM model in the training and testing stages is 0.32 and 0.78 respectively, and the root mean square error ( RMSE ) is 0.39 and
　　0.98 respectively, which is significantly better than the performance of the SVR model. The LSTM model can stably reproduce the seasonal warming and cooling laws of water temperature. The prediction results are highly consistent with the trend of the measured sequence, and maintain good convergence and consistency in multiple rounds of reconstruction. This study verifies the effectiveness of machine learning algorithm in the reconstruction of water temperature time series, and provides a feasible technical means for the problem of missing or discontinuous monitoring data of river water temperature.