臺灣位處西北太平洋颱風主要路徑上，近岸暴潮水位對颱風氣壓、風速、路徑及風場結構具有高度敏感性。傳統理想參數化風場模型具有計算效率高與資料需求低等優點，但多建立於軸對稱假設之上，較難描述侵臺颱風受中央山脈地形調制後所形成之非對稱氣壓場與 10 m 風速場。相較之下，高解析度大氣模式較能呈現颱風核心結構，以及臺灣地形調制下之非對稱風壓場特徵。基於此，本研究發展資料驅動式真實參數化風場模型（REalistic Parameterization model， REP），嘗試利用歷史大氣模式資料中所包含之地形調制風壓場特徵，作為降低即時氣象場建置成本之潛在方法。 本研究蒐集 1980 至 2023 年 ERA5 全球再分析資料，以及 1990 至 2021 年中央氣象署 TWRF 區域大氣模式資料，建立侵臺颱風歷史氣象場資料庫。REP 模型依據目標颱風之中心位置、中心氣壓、最大風速與最大風速半徑，搜尋具有相似位置、強度與尺度條件之歷史成員，並透過權重計算與加權合成產生海平面氣壓與 10 m 風速場。為改善合成場之中心定位與時間連續性，模型進一步加入拉格朗日平移修正與時間平滑處理。參數調定部分以 2016 年梅姬颱風為主要案例，最終選定 REP(TWRF) 採用 M20-L1-MA1，REP(ERA5) 採用 M10-L1-MA1 作為後續案例模擬設定。 本研究選取 2013 年蘇力颱風與 2016 年梅姬颱風進行案例分析，其中梅姬颱風用於 REP 參數敏感性分析與東部登陸過山型案例評估，蘇力颱風則用於檢驗北臺灣通過型颱風情境下之模型表現。為避免目標事件直接參與相似成員檢索而造成資料相依性，本研究於 REP 合成與參數敏感性分析過程中均採用目標颱風排除策略，將目標颱風之所有歷史成員自相似成員資料庫中排除，再進行類比檢索、權重計算與加權合成。後續比較 ERA5、TWRF、REP(ERA5)、REP(TWRF)、Holland（1980）理想風場模型（H80）及中央氣象署理想化風場模型（CWA）六組氣象驅動場，並將其輸入 COMCOT-SS 暴潮模式，搭配 TPXO 天文潮邊界條件，模擬臺灣沿岸暴潮水位偏差。評估項目包含二維氣壓場、10 m 風速場、暴潮水位偏差空間分布、測站時序，以及偏差（Bias）、相關係數、MAE 與 RMSE 等統計指標。 兩個案例結果顯示，ERA5 可呈現颱風大尺度環流與背景場特徵，但對近中心氣壓低值與局部強風極值之描述較為平滑；TWRF則因具較高時空解析度，在本研究案例中較能呈現颱風核心低壓、局部強風及地形調制後之非對稱結構。REP(ERA5) 大致延續 ERA5 之平滑大尺度風壓場特性；REP(TWRF) 則在多數氣象場與暴潮水位結果中較接近 TWRF 原始場，顯示高解析度資料庫有助於 REP 保留地形調制下之非對稱風壓場特徵。相較之下，H80 與 CWA 雖能快速建立颱風風壓場，但對中央山脈造成之環流破壞、背風側重組與局部非對稱水位響應之描述仍有一定限制。 整體而言，REP 模型可作為介於完整大氣動力模式與傳統理想參數化風場模型之間的氣象驅動場重建方法。其主要特色在於利用歷史大氣資料庫中已包含之地形調制資訊，重建可供 COMCOT-SS 使用之非對稱颱風風壓場，並於本研究案例中展現應用於臺灣颱風暴潮模擬與預警氣象驅動場建置之潛力。由於 REP 合成與參數敏感性分析過程均已排除目標颱風本身，案例結果可降低歷史資料庫重複使用所造成之資料相依性疑慮。然而，梅姬颱風同時作為參數調定與敏感性分析案例，因此其結果應解讀為模型設定過程中的案例評估。本研究之正式案例分析仍以蘇力與梅姬颱風為主，REP 於不同路徑與強度條件下之重建表現仍受到歷史樣本代表性、母體資料解析度及合成平滑作用影響，因此其適用範圍應依目標颱風條件與資料庫特性審慎判讀，作業化預警應用亦仍需進一步評估。;Taiwan is located along the major typhoon tracks of the western North Pacific， where nearshore storm-surge levels are highly sensitive to typhoon pressure， wind speed， track， and wind-field structure. Conventional idealized parametric wind models are computationally efficient and require relatively limited input data， making them useful for rapid storm-surge simulations and operational applications. However， these models are generally based on an axisymmetric assumption and therefore have difficulty representing the asymmetric sea-level pressure and 10 m wind fields induced by the topographic modulation of Taiwan’s Central Mountain Range. In contrast， high-resolution atmospheric models can more realistically describe typhoon inner-core structures， terrain shielding， circulation disruption， and lee-side reorganization. Nevertheless， their use in real-time forecasting or multi-scenario simulations may still be constrained by computational cost， update frequency， and data latency and availability. To address this issue， this study develops a data-driven REalistic Parameterization model (REP) that uses terrain-modulated wind-pressure structures embedded in historical atmospheric model datasets as a potential approach for reducing the cost of constructing atmospheric forcing fields for typhoon storm-surge simulations. This study collects ERA5 global reanalysis data from 1980 to 2023 and TWRF regional atmospheric model data from the Central Weather Administration from 1990 to 2021 to establish historical atmospheric databases for typhoons affecting Taiwan. The REP model searches for historical members with similar location， intensity， and size characteristics according to the target typhoon’s center position， central pressure， maximum wind speed， and radius of maximum wind. Sea-level pressure and 10 m wind fields are then generated through weight calculation and weighted compositing. To improve the center alignment and temporal continuity of the composited fields， Lagrangian translation correction and temporal smoothing are further applied. In the parameter calibration stage， Typhoon Megi (2016) is used as the main case for evaluating the sensitivity of the maximum number of composited members， temporal smoothing， and Lagrangian translation correction. The final settings adopted for subsequent case simulations are M20-L1-MA1 for REP(TWRF) and M10-L1-MA1 for REP(ERA5).
Typhoon Soulik (2013) and Typhoon Megi (2016) are selected for case evaluation， representing a northern Taiwan passing typhoon and an east-coast landfalling typhoon crossing the Central Mountain Range， respectively. Megi is used for REP parameter sensitivity analysis and for evaluating an east-coast landfalling and mountain-crossing case， whereas Soulik is used to examine model performance under a northern Taiwan passing-typhoon scenario. To avoid data dependence caused by direct reuse of the target event in the analog search， a target-storm exclusion strategy is applied during both REP compositing and parameter sensitivity analysis. All historical members associated with the target typhoon are excluded from the candidate database before analog searching， weight calculation， and weighted compositing are performed. Six atmospheric forcing datasets are compared: ERA5， TWRF， REP(ERA5)， REP(TWRF)， the Holland (1980) idealized wind model (H80)， and the idealized wind model of the Central Weather Administration (CWA). These wind-pressure fields are used to drive the COMCOT-SS storm-surge model， coupled with TPXO tidal boundary conditions， to simulate storm-surge residuals along the coast of Taiwan. The evaluation includes comparisons of two-dimensional sea-level pressure fields， 10 m wind fields， spatial distributions of storm-surge residuals， time series at buoy and tide-gauge stations， and statistical analyses using bias， correlation coefficient， mean absolute error (MAE)， and root mean square error (RMSE). The two case studies indicate that ERA5 can represent the large-scale circulation and background atmospheric features of typhoons， but tends to smooth the near-center pressure minimum and local extreme wind speeds. TWRF， owing to its higher spatial and temporal resolution， better resolves the typhoon low-pressure core， strong-wind regions， terrain shielding， and lee-side reorganization in the cases examined in this study. REP(ERA5) generally inherits the smoother large-scale wind-pressure characteristics of ERA5， whereas REP(TWRF) is closer to the original TWRF fields in most atmospheric and storm-surge results. This suggests that a high-resolution database helps REP retain terrain-modulated asymmetric wind-pressure structures. In comparison， although H80 and CWA can rapidly construct typhoon wind-pressure fields， they remain limited in describing circulation disruption， lee-side reorganization， and localized asymmetric surge responses caused by the Central Mountain Range. Overall， the REP model proposed in this study can serve as an atmospheric forcing reconstruction method between full atmospheric dynamical models and conventional idealized parametric wind models. Its main feature is the use of terrain-modulated information already embedded in historical atmospheric databases to reconstruct asymmetric typhoon wind-pressure fields for COMCOT-SS simulations. The case results suggest that REP， particularly REP(TWRF)， has potential for application in typhoon storm-surge simulation and atmospheric forcing construction for storm-surge forecasting in Taiwan. Because the target typhoon itself is excluded during both REP compositing and parameter sensitivity analysis， the case results reduce concerns regarding data dependence caused by direct reuse of historical database members. However， because Typhoon Megi is used for both parameter calibration and sensitivity analysis， its results should be interpreted as a case evaluation within the model-setting process. The formal case analyses in this study remain focused on typhoon Soulik and typhoon Megi. The reconstruction performance of REP under different track and intensity conditions is still influenced by the representativeness of historical samples， the resolution of the parent dataset， and the smoothing inherent in the compositing process. Therefore， its applicability should be interpreted with consideration of the target-typhoon conditions and database characteristics， while operational forecasting applications require further evaluation.
