探索基於機器學習之天氣類比法於預測極端降水事件之應用

Abstract

極端降水事件對防災減災構成重大挑戰，使得準確的預測更為重要。而天氣類比法利用歷史大氣模式尋找相似的過去事件，提供了一種具有物理一致性且經驗證的方法，在極端降水的預測中已被證明有效。然而，傳統的天氣類比方法在處理複雜的大氣狀態時，往往難以有效捕捉局部特徵與非線性關係。因此，本研究提出基於 AtmoDist 的機器學習方法，用於評估大氣相似性並提升降水事件的預測能力。

為了提高模型對局部特徵的敏感度，我們引入了空間加權處理與兩階段天氣類比技術，並將 AtmoDist 與傳統方法進行系統比較。此外，利用可解釋人工智慧 (XAI) 技術深入分析模型的決策過程，以揭示重要特徵與影響因素。

研究結果顯示，不同相似性計算方法對大氣變數的表現存在差異，反映變數特定評估的必要性。AtmoDist 模型在降水預測上優於傳統方法，尤其對局部降水事件的預測能力經空間加權後明顯提升。同時，結果也指出模型準確度未必能充分代表其預測能力，需進一步關注特徵重要性分析。透過 DeepSHAP 評估發現，AtmoDist 對具有高時空變異性的大氣變數表現出更強的關注。

本研究為改善降水預測模型提供了新的視角，並說明可解釋技術對理解模型行為和提升可信度的重要性。此外，結果凸顯機器學習方法在捕捉非線性特徵與空間變異性方面的優勢，為未來替代傳統方法提供了潛在應用價值。

Extreme precipitation events pose significant challenges for disaster prevention and mitigation, making accurate forecasting essential. Weather analog methods leverage historical atmospheric patterns to identify similar past events, offering a physically consistent and empirically validated approach that has proven effective in predicting extreme precipitation. However, traditional weather analog methods often struggle to effectively capture localized features and nonlinear relationships in complex atmospheric states. To address these challenges, this study proposes a machine learning-based approach using AtmoDist to assess atmospheric similarity and improve precipitation event prediction.

To enhance sensitivity to localized features, we incorporated spatial weighting techniques and a two-stage weather analog framework, systematically comparing AtmoDist with conventional methods. Furthermore, we applied explainable artificial intelligence (XAI) techniques to analyze the model's decision-making process, providing insights into key features and influencing factors.

Our findings reveal that different similarity metrics yield varying performances across atmospheric variables, highlighting the need for variable-specific assessments. The AtmoDist model outperforms traditional approaches in precipitation forecasting, particularly for localized precipitation events, with notable improvements after spatial weighting adjustments. However, results also suggest that accuracy alone may not fully capture predictive performance, emphasizing the importance of feature importance analysis. Using DeepSHAP, we observed that AtmoDist places greater emphasis on atmospheric variables with high spatiotemporal variability.

This study offers a new perspective on improving precipitation forecasting models and underscores the role of explainable techniques in enhancing model interpretability and reliability. Moreover, the results demonstrate the advantages of machine learning methods in capturing nonlinear patterns and spatial variability, highlighting their potential as viable alternatives to traditional approaches.