SeaGraph: A spatiotemporal onboard wave forecasting method

Traditional wave forecasting methods for ship navigation often struggle to model dynamic spatiotemporal processes because of data latency, limited resolution, and computational constraints. Existing deep learning approaches also face difficulties in predicting wave conditions ahead of ships, largely due to static modeling assumptions and their misalignment with time-varying navigational trajectories. To address these challenges, we propose SeaGraph, a spatiotemporal onboard significant wave height (SWH) forecasting framework that integrates multi-source temporal information from historical vessel trajectories with spatiotemporal features of target waypoints. Specifically, we design an adaptive spatiotemporal encoding mechanism to capture complex interactions and long-range dependencies among nodes. Experimental results demonstrate that SeaGraph consistently outperforms sequential baselines, including Long Short-Term Memory (LSTM) and Transformer, achieving the lowest errors in SWH forecasting, with a 1-hour-ahead Mean Absolute Error (MAE) of 0.0511 m and a Mean Squared Error (MSE) of 0.0086 m, corresponding to improvements of approximately 7.8% and 9.5% over Transformer, respectively. Moreover, under a 6-hour prediction horizon, SeaGraph limits error growth to 29.6%, indicating superior robustness for longer-term forecasting. These results suggest that SeaGraph effectively enhances spatiotemporal feature integration and adaptability, providing accurate wave forecasting support for ship navigation safety.