Predictive modeling study demonstrates reduced swell forecast errors using buoy-trained deep learning, indicating effective site-specific mitigation of numerical wave model bias.
Key Points
To develop a site-specific supervised deep learning framework that corrects systematic wave height and arrival time biases in numerical spectral wave forecasts.
Trained an attention-based encoder-decoder model on 3.5 years of historical ECMWF directional wave variance spectra matched to 1D variance density spectra from a deepwater moored buoy.
Conditioned the model with context variables, including local tidal water levels and temporal embeddings, over a 5-day forecast horizon.
Reduced swell significant wave height Root Mean Square Error from 0.16 to 0.12 m (a 25% reduction) and swell peak period error from 2.22 to 1.39 s (a 37% reduction) at a 120-hr lead time.
Mitigated regional premature swell arrival bias across a 5-day horizon, although the model had difficulty resolving stochastic variance in high-frequency wind seas.