A novel hybrid forecast framework for significant wave height based on VMD-WOA-BiLSTM

Accurate prediction of significant wave height (SWH) is challenging owing to the non-stationarity of ocean waves, particularly for long-term horizons where phase lag compromises forecast reliability. This study proposes a novel hybrid framework coupling Variational Mode Decomposition (VMD) and Whale Optimization Algorithm (WOA) with Bidirectional Long Short-Term Memory (BiLSTM) for 24-hour multi-step SWH forecasting. An auto-regressive input strategy is incorporated to eliminate time-delay errors. The model was developed using data from Atlantic Station 41004 and validated for spatial generalization on Stations 41013 and 41047. The proposed framework was systematically compared with five benchmarks, including Random Forest (RF) and single BiLSTM. Results demonstrated that the model achieved an overall R2 of 0.9900. Even at the challenging 24-hour horizon, the RMSE, MAPE, and R2 were 0.1155 m, 6.84%, and 0.9730, respectively, significantly outperforming BiLSTM ((Formula presented) ) and RF ((Formula presented) ). Furthermore, generalization analysis on unseen stations yielded high R2 scores exceeding 0.94. These findings confirm that the proposed framework possesses superior accuracy and spatial robustness, effectively mitigating phase lag for practical engineering applications.