台灣具有高密度地面氣象站，可即時提供地面氣壓、風向、風速以及溫、濕度觀測，為同化系統中提供關於中、小尺度環流如午後熱對流、梅雨鋒面與颱風等天氣系統的重要資訊來源。然而，近地表觀測值和模式模擬值仍存在顯著偏差，是地面資料同化的主要困難之一。地面觀測資料於對流尺度預報效益，在同化期間同時進行偏差修正成為需謹慎評估之重要技術。 本研究參考衛星資料同化上已發展成熟的變分偏差修正（Variational bias correction, VarBC）方法，推廣應用到地面觀測資料同化上，並應用於台灣氣象署的對流尺度作業預報系統（CWA RWRF）。VarBC利用多變數偏差模型（multiple-predictor bias model）中觀測增量（innovation）與預測變數之間的統計關係，達成在分析循環中自適性動態修正偏差。 針對2022年6月6日到6月8日梅雨鋒面事件和2019年7月22日至23日北台灣午後熱對流個案進行同化及預報實驗，在同化地面觀測10米風與2米溫溼度時加入了CWA提出的VarBC偏差修正模型。結果顯示，VarBC可減少分析場中的觀測減背景場（OMB）平均，降低了原先的水氣正偏差及風速偏差。在溫度偏差模型中，可由熱通量項掌握日夜變化相關偏差特徵，並由觀測與模式高度差異項模擬此高度落差而導致的溫度偏差。 模式驗證結果進一步顯示，VarBC不僅改善了分析的偏差，也對後續預報產生持續的正面影響，可提升模式的穩定性。觀測驗證部分，通過對地基GPS天頂向總延遲量（ZTD）和QPESUMS累積降雨的校驗，顯示VarBC能夠降低水氣濕偏差，並且可減少預報在小雨時段高估的降雨，尤其對長時間降雨個案效果較為顯著。整體而言，VarBC在提升地面觀測資料同化效益方面具有極大潛力。 ;Significant biases exist between near-surface observations and model simulation outputs, posing a major challenge in assimilating surface data effectively. This study extends the variational bias correction (VarBC) method, widely used in satellite data assimilation, to surface data assimilation with a convective-scale data assimilation system in Taiwan (CWA RWRF). Through VarBC, adaptive bias correction is achieved during assimilation by utilizing the statistical relationship between innovations and multiple predictors in the bias model. The impact of surface VarBC is examined with two cases: a Meiyu front event from 6 to 8 June 2022, and an afternoon thunderstorm event from 22 to 23 July 2019. The VarBC is applied to 10-meter wind, 2-meter temperature, and 2-meter humidity observations. Results indicate that the VarBC reduces mean observation-minus-background (OMB) values; particularly, it decreases positive moisture biases and wind speed biases. The bias model also captures temperature biases related to the diurnal cycle via a heat flux predictor and adjusts elevation-related temperature biases via a predictor of differences in observation and model elevation. Model verification confirms that the VarBC improves both the analysis and forecasts, enhancing model stability. In addition, verification against Zenith Total Delay (ZTD) and QPESUMS precipitation observations shows that the VarBC reduces moisture biases and mitigates precipitation overestimation during light rainfall periods, especially effective in long duration rainfall cases. Overall, the VarBC shows potential for advancing the assimilation of surface observations in Taiwan.
