本研究使用2公里解析度WRF模式，以WRF三維變分資料同化為基礎，選取2017年6月1至4日造成臺灣重大淹水之梅雨鋒面雨帶個案，評析同化雷達觀測對模式定量降水預報的影響。本研究設計三組實驗，分別為無雷達資料同化實驗（NoDA），冷啟動（Cold）和循環更新（Full）實驗。2017年6月1日18 UTC個案分析顯示，NoDA和Cold均有顯著模式水象粒子起轉（spin-up）的問題，導致第1小時預報的回波強度偏弱。NoDA實驗模式鋒面雨帶位置有偏北的預報誤差，這主要是全球模式初始場中鋒面位置誤差所致。在Cold實驗中僅同化回波和徑向風仍無法修正全球模式鋒面偏北的預報誤差。透過循環更新（Full）過程，高解析度雷達資料同化系統可有效改善模式預報之鋒面的位置，而且可以取得模式熱力、動力和水象粒子之間更好的平衡關係，從而減少模式的起轉問題。針對本個案累積連續55次的同化預報實驗，模式預報6小時累積雨量之FSS得分，校驗結果和個案類似，Full預報結果最佳，Cold次之，NoDA表現最差。特別在大於20 mm/6hr閾值以上的大雨區間，Full實驗的預報結果具有明顯的優勢。

This study investigates the impact of the radar data assimilation (DA) on the model quantitative precipitation forecast (QPF) by using the WRF 3DVAR. A total of three experiments for the severe Mei-Yu frontal rainfall event in 1-4 June 2017 were designed: no radar DA (NoDA), cold start (Cold), and full cycle run (Full). Case study shows that NoDA and Cold have apparent spin-up issue and result in the less rainfall forecast in the first one hour. The location of the model frontal rainband has the bias to the north in NoDA. It is because of the errors from the global model. Assimilating the reflective and radial wind in the Cold has similar bias as NoDA. However, the prediction of the frontal rainband in Full experiment outperformances the other two, showing the best QPF performance. In addition, the Full experiment also shows the less spin-up issue compared with the NoDA and Cold. The statistics over the 55 case also show the performance that the Full is the best, Cold the second, while NoDA the worst. In particular, the Full has significant QPF performance better than NoDA and Cold as the rainfall threshold is larger 20 mm/6hr.