查詢結果分析
來源資料
頁籤選單縮合
題 名 | 應用非監督學習與時間融合變換器優化水庫集水區中展期降雨預報=Applying Unsupervised Learning and Temporal Fusion Transformer Models to Optimize Medium and Extended-range Rainfall Forecast in Reservoir Catchment Areas |
---|---|
作 者 | 許家銓; | 書刊名 | 農業工程學報 |
卷 期 | 69:3 2023.09[民112.09] |
頁 次 | 頁57-71 |
分類號 | 328.63 |
關鍵詞 | 非監督式學習; 時間融合變換器; 中展期降雨預報; Unsupervised learning; Temporal fusion transformer model; Medium and extended-range rainfall forecast; |
語 文 | 中文(Chinese) |
DOI | 10.29974/JTAE.202309_69(3).0005 |
中文摘要 | 近年來,臺灣地區乾旱頻傳,水資源管理上對於有效降雨預報的需求日益增加;而隨著數值天氣預報技術精進,降雨預報的準確性和可靠性已有長足進展。本研究提出一種結合非監督式學習(Unsupervised learning)及時間融合變換器模型(Temporal fusion transformer model, TFT)的方法,用於優化中展期系集降雨預報之準確性。非監督式學習是一種機器學習的方法,無需事先進行資料標記,這大幅增加了資料處理效率及使用彈性,且可自動對輸入的資料進行分群。時間融合變換器模型是一種基於時間序列轉譯概念所發展的機器學習模式,能夠有效處理時序列資料,並捕捉其內在的關聯性和變化性,更能有效學習長範圍的時間依賴性,對於時間序列預測問題具有較高的效能。本研究以臺灣地區石門水庫、德基水庫及曾文水庫等3座水庫集水區為研究案例,以本研究提出之方法推估未來30日的逐旬累積降雨之預報,並與常見的機率擬合法平均(Probability-matched ensemble mean, PM)進行比較,在3處研究案例水庫集水區中,可降低約39 %至61 %的旬降雨預報誤差。藉由本研究所發展之優化架構,可提升降雨預報於水庫集水區的應用效能,為水資源管理單位理解雨量預報產品在水庫水源管理上的應用提供實用的工具。 |
英文摘要 | In recent years, Taiwan has frequently experienced droughts, leading to an increasing demand for effective rainfall forecasts in water resource management. Alongside advancements in numerical weather forecasting technology, significant progress has been made in the accuracy and reliability of rainfall forecasts. This study proposes a method that combines unsupervised learning and the Temporal Fusion Transformer model (TFT) to enhance the accuracy of medium and extended-range ensemble rainfall forecasts. Unsupervised learning, a method of machine learning that does not require prior data labeling, significantly improves data processing efficiency and flexibility, and it can automatically cluster the input data. The Temporal Fusion Transformer model is a machine learning model developed based on the concept of time-series translation, which effectively handles time-series data, captures its inherent correlations and variations, and efficiently learns long-range temporal dependencies, providing higher performance for time-series prediction problems. This study uses the cstchment areas of the Shimen, Techi, and Zengwen Reservoirs in Taiwan as case studies. The proposed method can estimate forecasts for accumulated rainfall over each ten-day period for the next 30 days, and the results were compared with the commonly used Probability-Matched Ensemble Mean (PM). In the three case study catchment areas, the method reduces ten-day rainfall forecast errors by approximately 39% to 61%. Through the optimization framework developed in this study, the application performance of rainfall forecasts in reservoir catchment areas can be enhanced. This study provides a practical tool for water resource management institutes to understand the application of rainfall forecast products in reservoir water management. |
本系統中英文摘要資訊取自各篇刊載內容。