Abstract
Accurate prediction of Significant Wave Height (SWH) is vital for marine engineering safety, yet balancing computational efficiency with physical consistency in long-sequence modeling remains a challenge for data-driven approaches. To address this, we propose the Physics-Guided Dynamic Graph Mamba Network (PG-DyMamba). By integrating oceanographic priors such as windwave relations, our Physics-Aware Graph Learner adaptively captures time-varying multivariate dependencies. Concurrently, the Mamba architecture processes long historical sequences with linear complexity. To ensure physical plausibility, the model employs a composite loss function based on energy conservation and fluid smoothness, effectively constraining predictions to adhere to fundamental physical laws. Empirical evaluations on Australia, NDBC, and North Sea datasets confirm that PG-DyMamba outperforms state-of-the-art baselines. Notably, the model achieves a 19.2% MSE reduction in 48-step prediction horizon on the Australia dataset, demonstrating its robustness for operational marine applications.