運用可解釋與深度學習模型的花蓮海岸浪高預測

Abstract

準確的浪高預測對海岸與海洋活動至關重要，尤其在台灣，冬季東北季風常帶來劇烈的浪高變化。本研究聚焦於花蓮地區，該地浪高資料呈現不規則的週期性與頻繁的短期弱訊號波動，使精準的浪高預測面臨挑戰。我們首先使用可解釋的機器學習模型，如隨機森林（Random Forest），深入探究輸入變數與浪高之間的關係，並找出關鍵的預測因子。接著在整合多個測站的觀測資料後，我們善用其地理與時空分布特性，大幅提升模型的準確性，尤以長期預測成效最為顯著。在此基礎上，我們導入先進的深度學習模型，包括 Transformer 和 Informer 架構，進一步強化預測能力。研究結果顯示，雖然引入鄰近測站資料能顯著提升準確度，但若模型過度依賴最近的本地浪高觀測值，則可能導致遲滯現象。雖然遠處站點的風速資料有助於長期的預測，但如何能有效減緩此遲滯現象仍是未來模型發展的重要課題。

Accurate wave height prediction is essential for coastal and oceanic activities, especially in Taiwan, where winter northeast monsoons cause substantial wave variations. This study investigates Hualien, where wave data shows irregular periodicity and frequent short-term fluctuations, complicating precise prediction. We first applied interpretable machine learning models, such as Random Forest, to examine relationships between input variables and wave height, identifying key predictive features. By integrating data from multiple stations, we leveraged spatiotemporal patterns to improve accuracy, especially over longer horizons. Building on these insights, we implemented deep learning architectures, including Transformer and Informer models, to enhance performance. Results show that surrounding station data improves accuracy, but reliance on recent local measurements introduces lag. While wind speed data from remote stations supports longer-term predictions, mitigating lag remains a key challenge for future modeling.