請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92893
標題: | 臺灣降尺度降雨預報之進階深度學習方法 An Advanced Deep Learning Rainfall Forecast Downscaling Method in Taiwan |
作者: | 張容慈 Rong-Cih Chang |
指導教授: | 曾琬鈴 Wan-Ling Tseng |
共同指導教授: | 陳柏孚 Buo-Fu Chen |
關鍵字: | 降尺度深度學習,地理注意力機制,尺度分離函數, deep learning downscaling,geographical attention,scale-separated loss function, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 降尺度的氣象預報提供了更精確和在地化的預測,這對臺灣複雜的地形和多樣的微氣候來說至關重要,因為大尺度的全球模型往往難以提供準確的天氣預報。本研究透過深度學習降尺度的方法,能更有效地進行災害管理並減輕極端天氣事件的影響。
本研究旨在透過利用深度學習技術,改善臺灣的降雨預報。本研究使用歐洲中期天氣預報中心(ECMWF)的數據,其傾向於低估大雨並高估小雨。為了減少這些不準確性,本研究採用了U-net和生成對抗網絡(GAN)進行降尺度。訓練和驗證使用了2021年ECMWF 9公里解析度的降水預報,並以中央氣象署的雷達降水資料作為標籤資料,性能評估則使用2022年資料作為測試集。 本研究根據所提出的模型在U-net架構新增地理注意力層(GAL),以有效捕捉臺灣獨特的地理特徵。此外,本研究提出創新的尺度分離損失函數(SPL),透過將降水數據分為大尺度平滑場和小尺度擾動場來優化模型。結果表明,GAN模型始終優於基線模型,而GAL模型在預測大雨事件表現出特別的優越性,後將這些模型綜合應用在DeepGAD模型中得到不同模型的綜合特性,提供更精細的降雨預報。其中,在SPL系列的測試中,DeepGAD(R+R')模型被認為是最準確的降尺度模型。本研究亦強調了案例分析結合關鍵成功指數分數的重要性,以全面評估每個模型的表現能力。另外,本研究使用NCEP-GFS 22公里解析度的資料做測試,結果還表明,這些模型未來可以用來修正或降尺度新興全球人工智慧模型的預報,如Pangu和FourcastNet,因為這些模型的解析度通常為25公里,與NCEP-GFS相近。最終,本研究透過應用深度學習方法顯著提高了臺灣降雨預報的準確性。 Weather forecast downscaling offers more precise and localized predictions, enhancing accuracy is critical for the intricacies of Taiwan's terrain and its diverse microclimates necessitate accurate weather forecasting, which large-scale models often struggle to provide. Through deep learning downscaling method, is more effective for disaster management and mitigating the impacts of typhoons and other extreme weather events. This research aims to enhance rainfall forecasting in Taiwan by leveraging deep learning techniques to address the region's challenging terrain. The study critiques the European Centre for Medium-Range Weather Forecasts (ECMWF) data, which has a tendency to underestimate heavy rainfall and overestimate light precipitation. To mitigate these inaccuracies, U-net and Generative Adversarial Networks (GAN) are employed for downscaling. Training and validation are conducted using 2021 ECMWF precipitation forecasts at a 9 km resolution, supplemented with labeled data from the Central Weather Bureau. Performance evaluation is carried out using a separate testing dataset from 2022. The proposed model integrates a geographical attention layer (GAL) within the U-net architecture to effectively capture Taiwan's unique geospatial characteristics. Additionally, a scale-separated loss function is implemented to optimize the models by dividing rainfall data into large-scale smoothing fields and small-scale disturbance fields. Findings indicate that the GAN model consistently outperforms the baseline, with the GAL model showing particular proficiency in predicting heavy rainfall events. The synthesis of these models into the DeepGAD model results in detailed rainfall distribution forecasts. The study also introduces the innovative concept of scale separation, with the DeepGAD(R+R') model identified as the most accurate downscaling model. The importance of case analysis, combined with critical success index (CSI) scores, is emphasized to fully assess each model's capabilities. The findings also imply that these models can be employed to correct or downscale forecasts generated by emerging global AI models, such as Pangu and FourcastNet, which typically operate at a 25-kilometer resolution, as evidenced by the NCEP-GFS test. Ultimately, this research significantly advances the accuracy of rainfall forecasting in Taiwan through the application of deep learning methodologies. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92893 |
DOI: | 10.6342/NTU202401379 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 氣候變遷與永續發展國際學位學程(含碩士班、博士班) |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-2.pdf 授權僅限NTU校內IP使用(校園外請利用VPN校外連線服務) | 6.79 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。