Tidal-level short-term forecasting is critical for disaster prevention, mitigation, and ecological management in nearshore areas. Traditional methods such as harmonic analysis and numerical simulations face limitations when handling large-scale tidal forecasting and real-time monitoring. This paper proposes a Temporal Convolutional Network (TCN) method for tidal forecasting, using tidal fields generated by physical hydrodynamic models as data sources. The method predicts the 1-h-ahead full-domain tidal field from historical sequences at four eastern boundary grid nodes. The TCN achieves higher accuracy than baselines including Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (Bi-LSTM), and k-nearest neighbors (KNN), providing empirical support for efficient coastal tide forecasting through the integration of physics-based simulation and deep learning.