An innovative deep learning model for accurate wave height predictions with enhanced performance for extreme waves

Enhancing wave prediction models, with a particular focus on accurately forecasting extreme waves, is crucial for advancing ocean engineering applications. Existing deep Learning models trained on continuous hindcast datasets often struggle to capture rare extreme events, resulting in reduced prediction accuracy. This study introduces the Extremum-Enhanced LSTM-NBEATS (EELN) model, a novel deep learning approach for highly accurate 24-h significant wave height (Hs) predictions, with a focus on extreme wave conditions. The EELN model integrates the Extremum-Enhanced process with the hybrid LSTM-NBEATS model, combining Long Short-Term Memory networks (LSTM) for extracting information from covariates and Neural Basis Expansion Analysis for Time Series Forecasting (NBEATS) for capturing temporal dependencies. The Extremum-Enhanced process dynamically selects models trained by extreme wave datasets over different thresholds, thereby minimizing prediction errors. The EELN model outperforms standalone LSTM and NBEATS models, achieving lower Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and higher coefficient of determination (R²) values. Its validation against buoy data highlights its robustness and reliability, particularly in predicting extreme wave events, demonstrating its potential as a valuable tool for forecasting significant wave heights in Ocean Engineering.