An explainable (interpretable) stacking ensemble machine learning model for real-time and short-term significant sea wave height prediction

• Developed an explainable wave height predictive model for marine measurement systems. • Established a stacking ensemble model for accurate wave height prediction. • Integration of SHAP for feature importance and model transparency analysis. • Analyzed feature contributions and interactions in the stacking framework. • Evaluated base learner and feature impacts on prediction performance. Accurate and interpretable wave height prediction is essential for advancing wave energy and marine renewable energy. This study proposes an explainable stacking ensemble machine learning framework for real-time and short-term significant wave height prediction. The framework constructs a diverse pool of nine candidate base learners from different algorithm families. Based on performance and correlation analysis, a representative subset is selected and aggregated using a linear regression meta-learner to enhance robustness. The framework is validated using 2016 observational data from NDBC buoys 41002 (North Atlantic) and 42055 (Gulf of Mexico). For real-time forecasting, the stacking ensemble consistently outperforms baseline models at both buoy sites, achieving the highest R 2 (0.9148 / 0.9391) and the lowest RMSE, MAE, and MSE values. It also maintains superior accuracy in short-term multi-step forecasting (1 h, 3 h, 6 h) compared to LSTM and TCN-LSTM, demonstrating strong robustness and generalization across varying time horizons. To ensure model transparency and practical trustworthiness, Shapley Additive Explanations (SHAP) are employed, enabling comprehensive interpretation of feature contributions and the relative influence of each base learner. By integrating SHAP with stacking ensemble learning, this study improves both the accuracy and interpretability of wave height prediction, supporting safer and more efficient use of marine resources.