本研究採用系集法(Ensemble-based method)，首次在臺灣藉由天氣個案之短期系集預報，針對中尺度天氣系統預報誤差結構進行分析。個案選取為2008年西南氣流聯合觀測實驗(SoWMEX IOP8)期間，於06月15至16日之間生成之中尺度對流系統。使用WRF區域模式3.2.1版，以72組系集預報結果作為樣本，運用統計方法估算出預報誤差協方差(Forecast Error Covariance)。在降水天氣系統期間，比較不同解析度(9-km與3-km)之誤差結構，顯示高解析度的方差量值較大，較能呈現中尺度對流系統在數值預報中之不確定性。同時，檢驗濕對流過程(Moist processes)對於誤差的影響，發現中高層風場以及溫度場之方差，與潛熱釋放作用有關。而近地表的風場受到摩擦效應使其方差受到抑制而減小。本研究同時探討時間與空間上的誤差相關性，動力相關變數影響範圍較大(約36公里以上)且時長較長(2 小時)，熱力相關變數則較為局地(約12公里左右)且在模式向後積分30分鐘後相關性遞減迅速。臺灣周邊之水平風場誤差特徵具有地形相依(Topography-dependent)結構，能夠提供臺灣地區進行觀測時之索引，以及評估資料同化的最佳策略，達到改善臺灣地區短期天氣預報之能力。

This study focuses on the short-term forecast error structures at convective scale. With a set of 72-member ensemble forecasts by Weather Research and Forecasting (WRF) model, the error covariances are presented. A case study during Southwest Monsoon Experiment intensive observing period 8 (SoWMEX-IOP8) in 2008 is investigated. The characteristics of forecast error covariances are examined by the variance and error correlation in state variables. Compared with different resolution, the variance of state variables are larger in higher resolution, particularly in rain mixing ratio. It indicates higher resolution run (3-km) can better-represent the smaller scale uncertainties in this severe weather event. The variance of horizontal wind and temperature at mid and high level are dominated by latent heat release. Additionally, frictional effect leads to near-surface wind variance suppressed. Moist processes not only impact on the distribution of variance, but paly important role of error correlation in temporal and spatial structure. The dynamic and thermodynamic variables indicate different performance of correlation length. The information of topography-dependent forecast error provides optimal strategies of data assimilation, especially for assimilating radar network over Taiwan area.