Answers are grounded in this paper and cite their sources.
Forecasts of moored-vessel motion help port operators identify safe operating windows, yet same-vessel training may overstate performance on unseen ships. We evaluate a variable-wise inverted transformer for one-hour-ahead prediction of significant motion amplitudes in all six degrees of freedom, using motion history, ocean-meteorological forcing, vessel geometry, and berthing information. The public dataset from the Outer Port of Punta Langosteira uniquely combines hourly amplitudes, concurrent forcing, vessel particulars, and a vessel-level partition; 81 vessels are used for development and 3 were held out for testing. Against persistence, linear regression, LSTM, DLinear, PatchTST, and a quasi-static baseline, the transformer attains the lowest mean absolute error for sway (0.063 m) and pitch (0.018°), while linear regression is best for roll and yaw; surge and heave (six windows each) are reported as case-specific observations. Among the four adequately sampled responses, the transformer ranks first in normalized mean absolute error (0.169), narrowly ahead of linear regression and persistence; the improvement is statistically supported for sway (p=0.035) but not for pitch (p=0.365). Error grows with lead time and wave height, CPU latency remains below 0.1 ms per window, and Monte Carlo dropout intervals are under-calibrated. Variable-wise attention offers a lightweight option on par with the strongest baselines, underscoring the strength of simple baselines and the importance of vessel-level holdout evaluation.
No takes yet. Share an insight, caveat, or question.
Cai et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: