Accurate short-term significant wave height forecasting is critical for coastal navigation, offshore energy operations, and hazard preparedness. However, operational wave forecast skill is often limited by uncertainty in wind forecasts. This study develops attention-based Long Short-Term Memory (LSTM) networks that explicitly address wind forecast uncertainty through noise-augmented training strategy and probabilistic modeling. Using wind and wave observations from the Western Long Island Sound (WLIS) buoy and operational North American Mesoscale (NAM) wind forecasts at Sikorsky Station, we characterize wind forecast errors and incorporate their statistical properties in model design. Three model variants are evaluated: (1) a perfect-future model trained with ideal future winds, (2) a noise-augmented model trained with controlled wind perturbations to enhance robustness, and (3) a probabilistic model that outputs both mean forecasts and uncertainty estimates. Results show that noise-augmented training reduced RMSE and MAE by approximately 7% compared to the perfect-future model and enhanced robustness to a broader range of wind forecast uncertainties. Furthermore, it led to smoother and more consistent performance across lead times. The probabilistic model achieved comparable accuracy while producing well-calibrated uncertainty estimates. Both models outperform a baseline relying solely on past buoy observations at all forecast horizons. These findings demonstrate that explicitly incorporating wind forecast uncertainty significantly improves wave height forecasting accuracy and reliability. The proposed framework offers a practical and robust pathway for operational wave forecasting systems under imperfect atmospheric forcing, enabling better-informed risk management and coastal decision-making. • Attention-based LSTM models using future winds outperform past-only baselines. • Noise-augmented training improves robustness to large wind forecast errors. • Probabilistic modeling provides calibrated uncertainty for risk-aware decisions. • Under forecast winds, proposed models reduce RMSE and MAE by about 7%.
Ren et al. (Mon,) studied this question.