Determine whether richer input spaces reveal architectural advantages

Determine whether multivariate input spaces containing variables such as wind, pressure, or spectral partitions reveal genuine performance advantages among DLinear, LSTM, PatchTST, ResAttLstm, and Mamba2 that are absent under univariate significant-wave-height input.

Background

The reported convergence of the five architectures is conditional on using only significant wave height as input, mean-squared-error evaluation, a per-buoy forecasting setup, and a specific log-transform and robust-scaling pipeline. The paper therefore does not establish that the architectures are equivalent under more informative physical representations.

The unresolved issue is whether differences between architecture families emerge when models receive atmospheric variables or spectral decompositions that expose forcing and wave-component information unavailable in the univariate formulation.