透過您的圖書館登入
IP:18.222.148.124
  • 學位論文

應用類神經網路於颱風降雨量即時預報之研究

The Research on Application of Artificial Neural Networks in Real-time Typhoon Rainfall Forecasting

指導教授 : 徐年盛
若您是本文的作者,可授權文章由華藝線上圖書館中協助推廣。

摘要


台灣由於位於西北太平洋之濱、颱風路徑之要衝,再加上河川坡陡流急、沿海地區地層下陷嚴重,颱風所帶來之洪水常致使人民生命財產損失;為了提供資訊予淹水預警與水庫防洪操作系統,短至長延時之颱風降雨預報為一相當重要之課題。颱風降雨為一混沌且不確定性高之系統,而調適性網路模糊推論系統(Adaptive Network-based Fuzzy Inference System)具有學習與模糊邏輯推理的能力,對於預測如降雨量這種非定律性颱風大氣-降雨關係效果可能較佳,因此本研究應用ANFIS並研發最佳化模式參數與架構之建構機制來建立提前1小時至6小時之颱風降雨預報模式,並與傳統最常使用之倒傳遞類神經網路(Back Propagation Neural Networks)做比較。本研究流程首先使用無母數之相關性分析來判別多種參數與降雨量之間的關係,決定最適合類神經網路之輸入變數。接著為了增進高雨量預報之精確度,本研究使用兩種架構來建構單一延時之雨量預報模式,第一種是單模式建立法,即直接建立一擁有最小誤差之模式;第二種是雙模式混合建立法,其方法為組合單一延時之高雨量與低雨量預報模式;而後比較單模式或雙模式混和兩種架構之優劣。而為了改善建構ANFIS雨量預報模式時搜尋最佳架構之效率與預報之準確性,本研究分別使用禁忌演算法與隨機試誤法搜尋減法聚類法之鄰近半徑,期望禁忌演算法能得到較隨機試誤法好之解。此外利用耦合之架構來建立短至長延時之雨量預報模式,期望能增進長延時雨量預報之精確度。最後比較BPN-耦合、BPN-非耦合、ANFIS-耦合、ANFIS-非耦合四種短至長延時預報模式架構來做比較。本研究以石門水庫集水區為研究區域,研究年限為西元2001至2009年,目的為預報玉峰以及霞雲雨量站提前1小時至提前6小時之雨量;結果顯示,禁忌演算法來優選ANFIS之最佳初始參數,並以雙模式混合法與耦合之架構來做短至長延時之颱風降雨預報,可得到最精確、快速且穩定之預報效果。

並列摘要


Taiwan is located in northwest side of the Pacific Ocean within the moving path of typhoons, and the slope of rivers is steep. Therefore the typhoons often cause a lot of damages of lives and properties. It is an important subject to provide information about typhoon rainfall forecasting for the flood warning and the flood control of reservoir operation system. The rainfalls in the typhoon period are uncertain and chaotic phenomena. Adaptive Network-based Fuzzy Inference System has the ability of learning and fuzzy logic reasoning, and it may be useful for the stochastic relationship between rainfalls forecasting and activities of the atmosphere. Therefore this study invents the best way to optimize ANFIS’s parameter and structures to forecast the typhoon rainfalls of 1 to 6 hours for future periods, and then the results are compared with the values of Back Propagation Neural Networks method. This study used a variety of correlation between parameters and rainfalls to determine the most appropriate input values of neural network. In order to improve the precision of rainfall forecasting, this study uses two methods: the first method is a single-model established with the smallest error; the second method is a dual-model contains higher and lower rainfall forecasting models, and then the superiority between these two models are discussed. In order to reform the efficiency and accuracy of ANFIS rainfall forecasting model, the tabu search and subtractive clustering method are used to determine the best ANFIS structure, and the solutions of tabu search method are expected better than the trial and error method. In addition, the coupling methods are embedded in the rainfall forecasting model to improve the accuracy of long-term forecasting. Finally, the frameworks of these four models, BPN coupling, BPN non-coupling, ANFIS coupling, ANFIS non-coupling with long-term and short-term forecasting are compared with each others. In this research, the study area are Shihmen Reservoir, and periods are from AD 2001 to 2009. To forecasting rainfall at Yufeng and Siayun rainfall stations ahead from 1 hour to 6 hours. Results show that combined with tabu search and subtractive clustering to optimal parameters ANFIS structure, dual-model and coupling methods for the typhoon rainfall forecasting , receives the most accurate, fastest and stable prediction results.

參考文獻


蔡孟蓉、江衍銘、張斐章,「類神經網路結合衛星影像於颱風降雨預報」,台灣水利,第五十五卷,第一期,第22−31頁,2007。
邱建堯,「結合GA 與CG 優選最佳倒傳遞類神經網路--以雨水下水道水位預測模式為例」,國立台灣大學生物環境系統工程學研究所碩士論文,2005。
行政院農委會水土保持局石門水庫集水區保育治理網站:
葉天降、樊彩虹、李昀寰,「台灣地區颱風降水之回歸預測(一)台北颱風降水線性回歸預測模式」,大氣科學,第二十九期,第一號,第77−96頁,2001。
葉天降、謝信良、吳石吉,「簡單統計方法於台灣地區颱風降水預測之研究:(二)預測結果隨區域之分布」,大氣科學,第二十八期,第四號,第263−279頁,2000。

延伸閱讀