The significant wave height ( \(\:{H}_{s}\) ) is a fundamental parameter governing wave energy assessment, offshore structural design, and marine hazard evaluation. Accurate prediction of \(\:{H}_{s}\) remains challenging because numerical wave models incur extremely high computational costs, largely due to the multiscale compound periodicity and strong randomness inherent in ocean wave systems. A hybrid deep learning framework, CBLA-XGBoost, is proposed for the accurate prediction of \(\:{H}_{s}\) , leveraging the complementary strengths of deep learning architectures and gradient boosting methods. Incorporating a Convolutional Neural Network (CNN), and a Bidirectional Long and Short-term Memory network (BiLSTM), and Attention Mechanism (AM), the framework is designed to extract deep features from historical data. Furthermore, the Extreme Gradient Boosting (XGBoost) algorithm is integrated via a multi-model fusion strategy based on weighted ensemble techniques. Input features are initially selected using Pearson’s correlation coefficient and the XGBoost feature importance score. Then, the selected features are fed into AM-based CNN-BiLSTM (CBLA) and XGBoost for training to predict \(\:{H}_{s}\) , and the outputs of these individual models are subsequently fused to generate the final predictions. The framework is evaluated for \(\:{H}_{s}\) forecasting at three stations in the North Pacific with 1-, 2-, 4-, and 6-hour lead times. Results demonstrate that CBLA-XGBoost outperforms the individual CBLA and XGBoost in terms of prediction accuracy and robustness, and significantly exceeds the performance of other benchmark models.
