在台灣夏季，弱綜觀條件下的主要降雨來源為午後對流事件，午後對流事件有著短延時強降雨的特性，本研究期望整合雷達資料與多來源氣象資料，利用機器學習演算法針對對流肇始及後續發展進行一小時即時預報，我們蒐集了 2017、2018、2019 年弱綜觀午後對流事件(共 33 天)的向日葵八號衛星、GPS 測站可降水量、氣象局地面測站以及QPESUMS 雷達回波，並利用機器學習演算法 Dense Neural Network (DNN)以及Convolutional LSTM Network (ConvLSTM)進行預報。結果顯示在 ConvLSTM 上採用 Focal loss α = 0.8, γ = 2有較好的結果，在2019測試資料集上點對點ETS分數為0.1181、鄉鎮市 ETS0.3569，和 Python 雷達迴波外延方法 Pysteps-OF 使用的光流法相比，ConvLSTM 在對流初始階段以及發展階段都有比較好的結果。本研究亦針對 20190524 以及 20190617 的案例，分析了 ConvLSTM 預報的特性，結果顯示在對流初始及消散期間，ConvLSTM 會有延遲預報的現象，而在對流發展期間則有最好的預報表現，本研究亦探討在雷達資料上加入不同變數資料的影響，結果顯示加入測站資料以及衛星資料，在對流初始的預報上都有更好的結果，顯示多元的氣象資料對於午後雷暴事件的預報是十分重要的。;In the summer of Taiwan,, the primary source of rainfall under weak synoptic conditions stems from Afternoon Thunderstorms (ATs) convective events characterized by short-duration heavy rainfall. This study aims to integrate radar data with various meteorological data to employ machine learning algorithms for real-time one-hour nowcasting of ATs convection. We collected data from the Himawari-8 satellite, precipitable water vapor from GPS stations, CWB surface stations, and QPESUMS radar echoes for weak synoptic ATs events (a total of 33 days) in 2017, 2018, and 2019. We utilized machine learning algorithms Dense Neural Network (DNN) and Convolutional LSTM Network (ConvLSTM) for forecasting. The results demonstrate superior performance when applying Focal loss with α=0.8, γ=2 in ConvLSTM.With a point-to-point ETS score of 0.1181 and a district ETS of 0.3569 on the 2019 test dataset,ConvLSTM outperforms the optical flow method used in Pysteps-OF, especially in the initial and development stages of convection. This study also examines the characteristics of ConvLSTM forecasts using case studies from May 24 and June 17, 2019. It was found that ConvLSTM tends to exhibit delayed forecasting during convective initiation and dissipation, yet it performs best during the development phase. Furthermore, the study investigates the effects of combining different variables with radar data. The results show that incorporating data from weather stations and satellite data significantly improves the initial forecast of convection. This highlights the importance of diverse meteorological data for forecasting ATs events.
