Site‐Specific Post‐Processing of Spectral Wave Forecast by Learning From Buoy Measurements

Abstract Operational wave prediction systems provide critical information across a range of marine industries and for public safety. Yet operational numerical wave forecasts can exhibit systematic biases, such as in wave height and arrival time of swell events. Focusing on a deepwater site in the Browse Basin off the northwest coast of Australia, here we described a method developed to mitigate these errors based on a site‐specific post‐processing supervised deep learning regression task. Using 3.5 years of historical spectral observations and numerical forecasts, we mapped the European Centre for Medium‐Range Weather Forecasts (ECMWF) directional wave variance spectrum to the one‐dimensional variance density spectrum recorded by a local moored wave buoy. To capture unresolved local hydrodynamics and seasonality, this forecast was augmented with specific context variables, namely local tidal water levels and temporal embeddings. The proposed architecture adapted an attention‐based encoder‐decoder model modulated with context‐conditioning. Results demonstrated that our model significantly improved upon the ECMWF baseline in the low‐frequency swell regime across a 5‐day forecast horizon. Specifically, at a 120‐hr forecast lead time, the model reduced the Root Mean Square Error for swell significant wave height from 0.16 to 0.12 m (a 25% reduction), and reduced swell peak period errors from 2.22 to 1.39 s (a 37% reduction). Furthermore, the architecture successfully mitigated the premature swell arrival bias characteristic of this geographic region. However, the network struggled to resolve the stochastic variance in higher‐frequency wind seas.