過去計算預報系統的誤差使用的是傳統觀測系統實驗的方法，近年多利用伴隨系統計算預報敏感度的方式來估計在短期預報中的觀測資料對誤差影響。本研究使用觀測資料的預報敏感度做為診斷工具來監測觀測資料在短期預報中的影響。 觀測資料來源主要使用福衛三號的真實掩星觀測(FORMOSAT-3/COSMIC)、福衛七號的模擬掩星觀測(FORMOSAT-7/COSMIC-2)和傳統地面站觀測資料(GTS)。對於2009莫拉克颱風以及2013海燕颱風分別進行12小時和24小時兩組實驗，其中針對掩星觀測做垂直分層分析，以互相比較福衛七號升空後與現行福衛三號之觀測改善程度的差異，且另外進行48小時預報作為參考。 伴隨模式的結果顯示，同化觀測資料對於颱風預報大部分具有正面的貢獻，在兩個颱風個案中， FORMOSAT-7/COSMIC-2表現較 FORMOSAT-3/COSMIC佳，對於擾動動能的正貢獻主要來自11000公里以上的高層。在使用完整物理參數化過程的48小時預報中，莫拉克颱風模擬結果顯示兩種掩星資料對於颱風預報並無太大分別，而在海燕颱風模擬中，FORMOSAT-7/COSMIC-2相比FORMOSAT-3/COSMIC路徑誤差可以再修正約10%左右。 伴隨模式雖然可同時提供多組觀測資料對於颱風預報的比較，但是其計算過程使用簡化過的模式，另外對於颱風來說，目標區域的選取是否能代表颱風駛流場、區域內海陸分布以及價值函數的選取等，都增加了結果的不確定性。 ;Observing System Simulation Experiments (OSSEs) are used to investigate current observational and data assimilation systems by testing the impact of new observations on them. Another robust tool has been developed as an adjoint-based model for assessing the impact of observations. This study use WRF-FSO system to understand their impact on short-range forecast error. The case chosen for the study are the typhoon Morakot (2009) and typhoon Haiyan (2013) with FORMOSAT-3/COSMIC GPS radio occultation refractivity data, simulated FORMOSAT-7/COSMIC-2 GPS RO data and GTS data. Two sets of experiments, 12-h and 24-h, have been designed to compare the RO data in vertical layers. We make a 48-h forecast for reference in addition. In the WRF-FSO experiment, it was found that FORMOST-7/COSMIC-2 is better than FORMOSAT-3/COSMIC data. The impact results indicate that larger impacts of GPS RO data are on upper levels for perturbation kinetic energy. In the 48-h forecast with full physics, Using FORMOST-7/COSMIC-2 RO data can improve about 10% in typhoon Haiyan track prediction than using FORMOST-3/COSMIC RO data. Although adjoint model can provide the Forecast sensitivity to observation at one time, but the measure still has some problem. First, it estimate the impact by simplified scheme and model. Second, we are not sure if the target zone can stand for the steering-flow region. Third, the land-sea distribution and cost function will affect the experiment result possibly. These problems increase the uncertainty of the result.
