使用深度神經網路預測弱綜觀環境下午後強對流及肇始–以WRF系集資料進行系統建置;Using Deep Neural Networks to predict afternoon thunderstorm and initiation under weak synoptic forcing - model development using WRF ensembles

NCU Institutional Repository > 地球科學學院 > 大氣物理研究所 > 博碩士論文 > Item 987654321/91719

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/91719

題名:	使用深度神經網路預測弱綜觀環境下午後強對流及肇始–以WRF系集資料進行系統建置;Using Deep Neural Networks to predict afternoon thunderstorm and initiation under weak synoptic forcing - model development using WRF ensembles
作者:	黃展皇;Huang, Chan-Huang
貢獻者:	大氣科學學系
關鍵詞:	對流肇始;深度神經網路;Convection Initiation;Deep Neural Network
日期:	2023-07-26
上傳時間:	2024-09-19 14:11:32 (UTC+8)
出版者:	國立中央大學
摘要:	台灣夏季在弱綜觀情況下，午後常有致災性且難以預報的強對流事件發生，因此快速且準確預測對流事件的肇始(Convection Initiation, CI)相當重要。由於小尺度強對流具有非線性特性，而深度神經網路(Deep Neural Networks, DNN)具備擬合非線性函數的能力，因此本研究針對弱綜觀強對流事件進行WRF (Weather Research and Forecasting)系集模擬，採用OSSE (Observing System Simulation Experiment)架構取得各式變數，並經過詳細的個案分析、預測目標定義與資料處理後再使用DNN進行強對流或CI的一小時預報，並且針對DNN進行多種如SHAP (SHapley Additive exPlanations)分析得知各變數的重要性，並依此展開許多面向的探討。本研究經過許多討論，可提供DNN系統參數等建置以及資料選用與處理的指南，且使用了許多不同面向的驗證方法，包括客觀分數、繪圖疊圖分析、性能圖、SHAP分析、persisFail實驗、坐標系測試實驗以及敏感度測試實驗，對模型性能、預測傾向以及變數重要性做了一系列討論並得出結果，變數重要性分析上與前人對於CI的統計結果接近，另外還考慮了提升DNN預報能力的方法例如加入空間相關的變數，以及在不同個案上應用的問題以及提出混合訓練的解決方式，最後預測表現雖無法與模糊邏輯做比較但顯著優於外延法，且僅需要以純測站變數組合即可快速應用於實際資料，若以中央氣象局標準將鄉鎮市行政區視為最小預報單位的情形，使用純測站可取得的資料進行變數組合，ETS (Equitable Threat Score)的預報得分對強對流為0.338，對CI為0.333，且ETS與事件發生頻率僅呈現低度正相關。 ;During Taiwan′s weak synoptic summer, there are often unpredictable and disastrous strong convective events occurring in the afternoon, making it important to rapidly and accurately predict the initiation of convection (CI). Due to the non-linear characteristics of small-scale strong convection, and the ability of Deep Neural Networks (DNN) to fit non-linear functions, this study focuses on weak synoptic strong convection events using the Weather Research and Forecasting (WRF) system and adopting an Observing System Simulation Experiment (OSSE) framework to obtain various variables. After detailed case analysis, prediction target definition, and data processing, DNN is used to forecast strong convection or CI for one hour, and various analyses like SHAP (SHapley Additive exPlanations) are conducted on DNN to determine the importance of each variable, leading to exploration in many different angle. After extensive discussions, this study provides guidelines for the establishment of DNN system parameters, data selection and processing, and uses various validation methods, including objective scores, graph overlay analysis, performance diagram, SHAP analysis, persisFail experiments, coordinate system testing experiments, and sensitivity testing experiments. A series of discussions were conducted on the model performance, prediction tendencies, and variable importance. The analysis of variable importance is consistent with previous statistical results for the CI, and methods to enhance DNN forecasting ability, such as adding spatially-related variables and solving problems with mixed training, were also considered. Although the prediction performance of the DNN cannot be compared with fuzzy logic, the results show a significant improvement over extrapolation methods, and can be quickly applied to actual data using only pure station variables. When using central weather bureau standards and considering the smallest forecast units as townships, the ETS (Equitable Threat Score) forecast scores for strong convection and the CI are 0.338 and 0.333, respectively, and the ETS and event occurrence frequency only weak positive correlation.
顯示於類別:	[大氣物理研究所 ] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	20	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....