A two channel optimized SWH deep learning forecast model coupled with dimensionality reduction scheme and attention mechanism

Abstract

In recent years, significant wave height (SWH) prediction based on deep learning has become a research hotspot. Input of related meteorological factors and time-frequency decomposition technology can effectively improve the SWH prediction accuracy. But at the same time, it is prone to cause dimensional catastrophe. Considering different characteristics, two dimensionality reduction schemes adapted to the related meteorological factors and time-frequency decomposed components are presented, which can effectively reduce the input dimensionality by about 70 %. A frequency-aware two-channel architecture that utilizes permutation entropy to classify components into high-frequency and low-frequency groups, achieving 60 % improvement in prediction accuracy (minimum mean absolute error (MAE) of two-channel model is about 0.01). Through the integration of Bayesian optimization and attention mechanisms, our optimized framework delivers a substantial 35 % increase in prediction accuracy. The proposed model maintains high prediction accuracy even under extreme wave conditions. Specifically, for SWH values exceeding 4 m, the model achieves MAE of less than 0.04 in 1-h-ahead prediction, demonstrating its robustness in challenging scenarios.