在這篇研究論文中，提出了一個一小時的天氣類別預測系統。主要的貢獻在於提出了一個以臺灣地區多維度原始資料，能夠短期的預測與分辨地區天氣狀態的系統。並提出對於氣象測站、環流場資料以及日本向日葵八號衛星的紅外線通道這三種不同之多維度原始資料做前處理、以及訓練之方式。且由於考慮到每個地區不同的氣候，提出對每個區域以測站為基礎分別建模的方式訓練模型。多維度的資料可以提供同一個目標的不同面向給模型，使其在學習時，較能以更為立體、不同面向的視角來看同一個天氣事件。 並在與現存研究中提出的天氣分類預測方法比較中都獲得更好的表現。並且在一小時是否降雨的預測下，與區域數值天氣預測模型WRF以及目前最好的降雨預測方法之一的ExAMP比較下皆獲得較好的表現。此系統可以在未來可能可以做為智慧城市的元件作使用，讓使用的民眾能預防即將到來的天氣。 ;In this research paper, a one-hour weather classification prediction system is proposed. The primary contribution lies in the development of a system capable of short-term prediction and classification of local weather conditions using multi-dimensional raw data from Taiwan. The system involves preprocessing and training with three different types of multi-dimensional raw data: meteorological station data, circulation field data, and infrared channel data from Japan′s Himawari-8 satellite. Considering the diverse climates in different regions, the approach involves training models separately for each area based on station-specific data. The use of multi-dimensional data provides various perspectives on the same target, enabling the model to learn and interpret weather events in a more comprehensive and multi-faceted manner. The proposed system demonstrates superior performance compared to existing weather classification prediction methods. Additionally, in one-hour precipitation prediction, the system outperforms both the regional numerical weather prediction model WRF and ExAMP, one of the best current precipitation prediction methods. This system has the potential to serve as a component of smart cities in the future, allowing residents to anticipate and prepare for imminent weather conditions.
