On the Limits of Univariate Deep Learning for Significant Wave Height ForecastingNew

The paper conducts a systematic hyperparameter search over DLinear, LSTM, PatchTST, ResAttLstm, and Mamba2, using nine context lengths from 1 to 168 hours for forecasting at NDBC buoy 41009. It then re-evaluates the best configurations on a corpus covering 47 buoys over 37 years, where the five model families reach a common performance level with a between-family standard deviation of 0.0014 square meters, or 0.8% of the grand mean.