Abstract: Accurate prediction of water level in waterway is of great significance for ensuring ships’ navigational safety. The lower Jingjiang waterway in the Yangtze River was taken as the research area, and the hydrological data from 2019 to 2020 and from 2021 were adopted as the train set and test set, respectively. A temporal convolution network (TCN) model was developed for water level prediction of the Lower Jingjiang waterway. Then the long short-term memory network (LSTM) and the support vector machine (SVM) were constructed for accuracy comparing with TCN. The results showed that there were differences of optimal input time windows of TCN in different stations. Jianli station, Tiaoxiankou station and Shishou station’s optimal input time windows were 2 days, 2 days and 3 days, respectively. In 2021, the Nash-Sutcliffe efficiency coefficient and determination coefficient of the water level prediction of TCN at each station in the lower Jingjiang River were higher than 0.995, and the RMSE was basically below 0.21 m. The overall performance of TCN was better than LSTM, and both of them could accurately predict the water level process and perform better than SVM. However, with the increase of prediction time scale, the prediction accuracy of water level showed a downward trend. In terms of different periods, the absolute error of TCN water level prediction in dry season was basically below 0.2 m, indicating that TCN has a great potential in the field of water level prediction. By analyzing the applicability and superiority of the TCN model, this study can provide technical support for improving the accuracy of water level prediction in the Yangtze River channel and the safe navigation of ships.