熱帶氣旋（TC）路徑預測對科學家與研究人員而言仍是一項挑戰；研究人員持續改進預報，以支援更準確的災害預警、疏散規劃與減災措施。熱帶氣旋對西北太平洋（WNP）地區影響顯著，但藉助人工智慧（AI）天氣預報模式（例如Pangu-Weather），可以產生數據驅動的預報，並有效地模擬複雜的大氣行為。本研究旨在透過引入Pangu-Weather目前尚未實現的預報校正和改進技術來提高熱帶氣旋路徑預測的準確性。 本研究採用四步驟流程。首先，從ERA5再分析資料產生輸入數據，並將其轉換為NumPy陣列。接著，將 Pangu-Weather 作為預訓練的確定性大氣預報核心，以 6 小時間隔產生未來 7 天的大氣預報場；Pangu-Weather 本身並不直接輸出熱帶氣旋路徑。之後，利用 TempestExtremes（TE）從這些大氣預報場中擷取初始熱帶氣旋預報位置，再透過深層平均導引、平均海平面氣壓異常中心重新定位、厚度導引，以及 TE 錨定平滑與護欄方法進行後處理，以調整預報路徑的經緯度位置。最終產生確定性的 TC 路徑預報，並以 IBTrACS 作為參考資料，利用直接位置誤差（DPE）、均方根誤差（RMSE）及平均絕對誤差（MAE）評估原始路徑與後處理路徑的表現。 本研究對2023年至2024年的43個TC進行了評估，結果顯示後處理總體上有所改善，但效果並不均勻。結果顯示，平均RMSE從478.74 km降至469.64 km，平均MAE從387.09 km降至380.47 km。在43個熱帶氣旋中，21個的均方根誤差（RMSE）有所改善，14個保持不變，8個有所惡化。例如，Ewiniar、Ampil、Bolaven和Soulik等熱帶氣旋的誤差降低幅度較大，而其他熱帶氣旋的改善幅度較小或略有惡化。這些結果表明，所提出的框架最好被理解為一種選擇性的改進層，可以改善許多合理的AI預測路徑，但並非通用的校正方案。;Tropical cyclone (TC) track prediction remains a challenge for scientists and researchers， who continually refine forecasts to support accurate disaster warnings， evacuation planning， and mitigation efforts. TCs significantly affect the Western North Pacific (WNP) region， but with the help of artificial intelligence (AI) weather forecasting models， such as Pangu-Weather， data-driven forecasts can be generated and can represent complex atmospheric behavior efficiently. This study seeks to improve TC track prediction by incorporating forecast correction and refinement techniques that are not yet implemented in Pangu-Weather. For this study， a four-step process is conducted. First， the inputs are generated from ERA5 reanalysis data and converted to NumPy arrays. Next， Pangu-Weather is used as a pretrained deterministic forecasting core to generate atmospheric forecast fields at 6-hour intervals over a 7-day horizon; it does not directly generate the TC track. TempestExtremes (TE) is then used to extract the initial TC forecast positions from these atmospheric fields， after which post-processing is applied using deep-layer-mean steering， mean sea level pressure anomaly recentering， thickness steering， and TE-anchored smoothing and guardrails. Finally， a deterministic TC track forecast is produced. The resulting tracks were evaluated against IBTrACS reference data using Direct Position Error (DPE)， Root Mean Squared Error (RMSE)， and Mean Absolute Error (MAE). A total of 43 TCs from 2023 to 2024 were evaluated and showed a net-positive but non-uniform response to post-processing. The results showed that the mean RMSE decreased from 478.74 km to 469.64 km， and the mean MAE decreased from 387.09 km to 380.47 km. Of the 43 TCs， 21 improved in RMSE， 14 remained neutral， and 8 degraded. Stronger reductions in errors were observed in cases such as Ewiniar， Ampil， Bolaven， and Soulik， whereas other storms showed weaker improvements or slight degradation. These findings indicate that the proposed framework is best interpreted as a selective refinement layer that can improve many plausible AI-derived tracks， but not as a universal correction scheme.
