Optimizing significant wave height forecasting through Ensemble Patch-TST and attention-enhanced recurrent models

Abstract

Accurate significant wave height (SWH) forecasting is pivotal for maritime safety, coastal engineering, and offshore operations. This study explores three advanced deep learning models: Ensemble Patch-TST, LSTM with Attention Mechanism (SWH-AT-LSTM), and GRU with Attention Mechanism (SWH-AT-GRU) for SWH forecasting across multiple forecast horizons. The innovative Ensemble Patch-TST model combines the strengths of ensemble learning and transfer learning to achieve state-of-the-art performance, particularly in medium- and long-term predictions. Using datasets from four distinct offshore stations, the study evaluates each model's ability to capture temporal dependencies and generalize across diverse environmental conditions. Results demonstrate that the Ensemble Patch-TST outperforms baseline Patch-TST and other state-of-the-art models, achieving superior accuracy and robustness, particularly for datasets with limited training samples. This work not only highlights the transformative potential of advanced deep learning techniques in enhancing maritime forecasting systems but also emphasizes their scalability and efficiency, providing a practical and effective approach for real-world applications.