結合時空克里金與Transformer神經網路架構於臺灣時降雨量預測模型之建構

Abstract

本研究旨在建立融合時空資料補值與深度學習模型的逐時降雨預測模型，針對臺灣地區氣象資料進行系統性分析與最佳化。研究以有人氣象站資料為黃金標準，透過時空克里金法補值自動站資料，並依縣市分別建構模型以掌握區域特性。模型比較結果顯示，Transformer（視窗長度=1）具最佳表現，遂作為基準模型，並進一步加入空間平滑項（Spatial Smoothing Term）與卷積神經網路（CNN）進行特徵擷取，並移除氣壓變數以增加模型的穩定性。最後，透過Softplus函數轉換以解決負降雨值問題，建構出最終模型「STformer」。實驗結果顯示STformer 預測模型顯著提升雨量預測準確度及物理合理性，並展現穩定且具泛化能力的預測表現，具高度實務應用潛力。

This study developed an hourly rainfall prediction model for Taiwan, integrating spatio-temporal data imputation with deep learning techniques. We used data from manned weather stations as a gold standard, applying spatio-temporal Kriging to impute missing automatic station data. To better understand and adapt to Taiwan's diverse weather patterns, we built a unique prediction model for each county. Through extensive model comparisons, a Transformer architecture with a window length of one consistently showed the best performance, making it our chosen baseline model. We further refined this by adding a Spatial Smoothing Term and employing Convolutional Neural Networks (CNNs) for robust feature extraction. To enhance model stability, the atmospheric pressure variable was removed. Finally, we incorporated a Softplus function to address the issue of negative rainfall predictions, resulting in our final model, called "STformer." Experimental results confirm that STformer markedly improves rainfall prediction accuracy and physical consistency. Its stable and generalizable predictive capabilities underscore its high potential for real-world applications in rainfall forecasting.