A data-driven water level prediction model for the lower reach of the Yangtze River is developed based on the deep learning Transformer method. The forecasting accuracy of this model is analyzed at stations along the tidal reach of the Yangtze River and then compared with other methods, including the Non-Stationary Harmonic Analysis (NS_TIDE) model,the combined Non-Stationary Harmonic Analysis with Auto-regressive model (NS_TIDE+AR) and hydrodynamic model. Statistical results demonstrate the effectiveness of the Transformer model for water level prediction in this region. For 24-h predictions, the Transformer model achieves RMSE values of 0.07–0.11 m along the Yangtze Estuary, outperforming the NS_TIDE method (0.20–0.23 m) and the NS_TIDE+AR method (0.08–0.14 m). The Transformer model can accurately predict water levels during the flood season but is unsuitable for storm surge forecasting due to the lack of more meteorological parameters, such as typhoon track, typhoon intensity as additional inputs.