深度學習模式對西北太平洋熱帶氣旋事件的展期預報能力

Abstract

隨著過去兩年來 AI 模型的快速發展，越來越多的深度學習天氣預報（Data-Driven Weather Prediction, DWP）模型不斷出現，能夠提供快速且方便的預測。在颱風預報方面，從過去的案例中可以看出，AI 模型在許多情況下並不遜於傳統數值模式，甚至能超越之。本研究針對2023年西北太平洋颱風個案，利用DWP模型在不同預測時效第一週（W1）至第四週（W4）所產生的系集預報，測試並比較不同 DWP 模型對颱風路徑的預測能力，並進一步了解其預報特性。在集合颱風路徑評分方法中，我們採用了 LHV（Likelihood Value）評分方式，並定義了一種創新的訊號偵測方法以檢測正確預報（Hit）、缺報（Miss）、誤報（False Alarm）、正確缺報（Correct Negative）。比較兩個 AI 模型（Pangu 與 FCNV2）在所有個案中的整體預報表現後發現，Pangu 模型在 W4 預報時較容易出現颱風缺報，而 FCNV2 則有較高的誤報傾向。從大尺度環境的預報能力來看，Pangu 模型在 W4 之後會低估 Z500 強度，而 FCNV2 則能維持其強度。此外，Pangu 模型在 W3 之後逐漸失去季風槽結構，可能導致其產生颱風的能力下降，因此產生的颱風數量較少，且分布在較高緯度地區。相較之下，FCNV2 能夠維持季風槽結構，導致產生較多颱風，並集中在低緯度的合流區域。透過延伸實驗，我們也確認了季風槽在擾動成長中的關鍵作用。最後使用 Wheeler-Kiladis 圖，我們比較了 ERA5、Pangu 預報與 FCNV2 預報的波動特性。結果顯示，在我們實驗目標的2023年7-10月的四個月間，FCNV2 與 ERA5 更為接近，但兩模式在東傳波動存在顯著差異，這些差異為未來的進一步研究提供了機會；而在較長時間的預報結果中，兩模式之Wheeler-Kiladis 圖較為相近。在整體模式預報波動的趨勢是否有顯著的差異造成大尺度結構的不同，可能需要未來更長的預報時間以驗證。

With the rapid development of AI models over the past two years, an increasing number of Data-Driven Weather Prediction (DWP) models have emerged, offering fast and convenient forecasting capabilities. In terms of typhoon cyclone forecasting, past cases have shown that AI models are, in many instances, not inferior to traditional numerical models and can even outperform them.In this study, we utilized ensemble forecasts generated by DWP models at different lead times (W1 to W4) for typhoon cases in the Northwest Pacific in 2023 to test and compare the ability of various DWP models to predict typhoon tracks, and to further understand the characteristics of these forecasts. For ensemble typhoon track evaluation, we adopted the LHV (Likelihood Value) scoring method and introduced an innovative signal detection framework, which includes Hit, Miss, Correct Rejection, and False Alarm.By comparing the overall forecasting performance of two AI models, Pangu and FCNV2, across all cases, we found that the Pangu model exhibited a higher rate of typhoon misses (Miss) at W4, while FCNV2 showed a greater tendency for false alarms (False Alarm). In terms of large-scale environmental prediction, the Pangu model tended to underestimate the intensity of Z500 after W4, whereas FCNV2 was able to maintain its intensity. Additionally, the Pangu model gradually lost the monsoon trough structure after W3, potentially contributing to a reduction in its typhoon generation capability. As a result, the number of typhoons it generated was smaller and tended to be scattered at higher latitudes. In contrast, FCNV2 was able to maintain the monsoon trough, resulting in a greater number of typhoons concentrated in the low-latitude convergence region.Through experiments, we also confirmed the crucial role of the monsoon trough in disturbance growth.Finally, using the Wheeler-Kiladis diagram, we compared the wave characteristics of ERA5, Pangu forecasts, and FCNV2 forecasts. The results showed that during our targeted four-month period (July to October 2023), FCNV2 was more similar to ERA5, though there were still significant differences in eastward-propagating waves between the two AI models. These differences present opportunities for further research. Over the longer period, the Wheeler-Kiladis diagrams of the two models were more similar. Whether these overall differences in model-predicted wave characteristics lead to significant variations in large-scale structures may require longer forecast periods in future studies to verify.