Attention-based ensemble learning for hindcasting significant wave heights and estimating extreme return levels

• A novel ensemble machine learning framework with attention-enhanced LSTMs improves prediction of both general and extreme wave heights. • The stacked model integrates specialized regressors and a classifier via a gating network, achieving a year-round RMSE of 0.711 feet (4.4% lower than RegGARCH). • A 31-year hindcast (1973–2004) enables high-resolution wave height reconstruction for extreme value analysis. • Incorporating hindcast data increases the 100-year return level from 11 feet to over 17 feet, while also narrowing statistical uncertainty. • The framework highlights the risk of underestimating extremes from short observational records and establishes a statistically robust foundation for coastal hazard assessment. Accurate prediction of significant wave heights is critical for coastal risk management, particularly under extreme weather conditions. This study presents an ensemble machine learning (ML) framework that integrates multiple attention-enhanced Long Short-Term Memory (LSTM) models to improve both general and extreme wave height prediction. The framework consists of three components: a general regressor, a big-wave-focused regressor with wave-weighted training, and a classifier for extreme wave detection. A stacked model integrates their outputs through a gating network, enabling dynamic adaptation across seasonal regimes. Compared to individual base learners and a statistical RegGARCH benchmark, the ensemble achieved improved accuracy across different wave conditions. Applied to hindcasting over 1973–2004, the framework produced a high-resolution dataset that enhanced extreme value analysis (EVA), increasing the estimated 100-year return level from approximately 11 feet (using observations only) to over 17 feet, with narrower confidence intervals. These results highlight the potential underestimation of extremes when relying solely on limited observational records. The integration of attention-based LSTM models, ensemble learning, and long-term hindcasting offers a data-driven approach to improve the reliability of wave height predictions and support more informed coastal hazard assessment.