Accurate water level forecasting is essential for coastal hazard mitigation, navigation safety, and estuarine management. However, the nonlinear interactions among astronomical tides, river discharge, meteorological forcing, and complex estuarine topography present significant challenges for reliable prediction. This study proposes a Temporal Graph Neural Network (TemporalGNN) for 24-hour water level forecasting in the Yangtze Estuary. The model operates directly on the unstructured FVCOM mesh and integrates multiple physical forcing variables, including wind fields, atmospheric pressure, upstream river discharge, open-boundary tidal elevation, and major tidal constituents. Temporal convolution, graph-based spatial learning, and recurrent temporal modeling are combined to capture the complex spatiotemporal dynamics of estuarine water levels. The model was trained and evaluated using a one-year high-resolution FVCOM dataset containing 28,111 computational nodes. Experimental results show that the proposed model achieves a global root mean square error (RMSE) of approximately 0.16 m with an overall R² of 0.9767 for 24-hour forecasts. During storm surge events, the model reproduces meteorologically induced water-level variations more accurately than conventional astronomical tide prediction. The proposed framework provides an efficient AI-based surrogate model for accurate and computationally efficient estuarine water level forecasting.