Wave height prediction: five deep learning architectures finish neck and neck; the buoy matters more than the model

Across 47 buoys and 37 years of measurements, LSTM, PatchTST, Mamba2 and the like are all on a par, and perform worse than simple persistence for the top 1 per cent of the most extreme waves.

The study screens five families of models — DLinear, LSTM, PatchTST, ResAttLstm and Mamba2 — and nine historical lengths, from 1 to 168 hours, to predict the significant wave height from the single series measured by a buoy. The settings are optimized on buoy 41009 of the American NDBC network, then re-evaluated on a corpus of 47 buoys covering 37 years.

The verdict: on all the buoys, the five architectures converge towards the same level, with a difference between families of 0.8% of the average error, while the error varies by a factor of 4.83 from one set of buoys to another. All of them beat persistence (which consists of predicting the last value observed), but none stands out. On a single buoy, the gain peaks between 12 and 24 hours, and the networks do worse than the persistence on the 1% of the strongest waves.

For the authors, architectural engineering has reached its limits with a single input variable: progress will come from wind and atmospheric pressure, from the transfer between buoys and from the separation between swell and wind sea.