Abstract: Accurate hydrologic time series prediction of water level and water volume is an important basis for water resources management and plays an important role in water transfer detection.Affected by factors such as the flow and water level of upstream tributaries, traditional single-factor water level prediction models cannot effectively consider these factors and water level prediction of Xijiang River faces severe challenges.Taking the typical Wuzhou Station on Xijiang River mainstream as the research object, a multi-factor water level prediction model based on splice-LSTM is established.The spliced Long Short-term Memory network(LSTM) and the fully connected linear model(Linear) were used to analyze and predict the flow and water level of the Xijiang River mainstream like Wuzhou station and other stations from 2020 to 2021.Research results show that:(1)The splice-LSTM can link to a non-linear layer and thus increase the weight of recent historical input data, making the model prediction closer to the historical value and reducing the prediction error.The linear part can improve the sensitivity of the model to linear components and the model to linear components, making the model′s prediction at the water level peak more accurate.(2)Compared with the traditional single factor ARIMA model and LSTM model, the accuracy of the split-LSTM model in water level prediction has increased by 14.4% and 10.1% respectively.The research results can provide a scientific reference for the precise pre-scheduling of ships by the Xijiang Shiplock Operation and Dispatching Center.