張斐章臺灣大學：生物環境系統工程學研究所丁裕峰Ting, Yu-FengYu-FengTing2007-11-272018-06-292007-11-272018-06-292004http://ntur.lib.ntu.edu.tw//handle/246246/56115水文事件的推估工作一直是從事水文工作者的一大工作課題；類神經網路為一黑盒分析之系統模式，目前已廣泛應用於各類水文工作之推估與分析並且都可獲得相當不錯的表現與評價。然而由於各式類神經網路所具有之特色與適用的資料性質、範圍皆有相當程度之差異性與不確定性；對於不同水文事件之真實物理機制與複雜，單一種類與架構的類神經網路雖可於水文事件之模擬工作獲得良好的效能表現，但是在實際應用時仍不難發現其難以描述及掌握不同水文事件之特性，仍至造成誤判或偏差。 本研究提出一新型併聯式類神經網路架構，並以分類式非監督式學習的自組織映射圖類神經網路(Self-Organizing Map，SOM) 為基礎，與監督式學習的共軛梯度倒傳遞類神經網路(Conjugate Gradient Back Propagation，CGBP )進行比較，期望能探討不同類神經網路對於不同水文事件之合適性，從而提出結合不同網路架構的併聯式類神經網路以改進其推估不同水文事件之能力，解決傳統單一架構類神經網路之缺點。本研究以蘭陽溪流域下游之蘭陽大橋流量站為研究對象，驗證併聯式類神經網路、與單一類神經網路(如：SOM、CGBP)對於蘭陽溪流域暴雨事件的洪流量推估效能；由結果的比較可以證實併聯式類神經網路有最好的效能表現。Accurate predicting hydrological event’s variation remains one of the most important and challenging tasks. Artificial neural network (ANN) is described as an information process system that consists of many nonlinear and densely interconnected processing units. With this parallel-distributed processing architecture, ANN has proven to be an efficient way for hydrological modeling and widely used for flood forecasting. Each neural network has its own character and suitability for different data structures. For modeling a complex physical mechanisms such as watershed rainfall-runoff process, a single architecture of artificial neural network is hard, if not impossible, to make good descriptions for different hydrological characters and to maintain highly applicable forecasting system. This research proposes an Incorporated Artificial Neural Network (IANN), which combines two ANN architectures for modeling the rainfall-runoff processes. The proposed method is compared with two famous ANN, the unsupervised learning Self-Organizing Map (SOM) and supervised learning Conjugate Gradient Back-Propagation (CGBP), by using Lan-Yan River data sets. Our goal is to find out the most suitable ANN for different hydrological events and to integrate the advantage of different kinds of ANN for improving the flood forecast ability. To demonstrate the applicability and capability of the proposed IANN structure, the Lan-Yan river, Taiwan, was used as a case study. For the purpose of comparison, three kinds of ANNs (SOM, BP, and the incorporated ANN) were performed. The results show that Incorporated Artificial Neural Network has superior performance than any other single artificial neural networks and can accurately make an one-step and two-step ahead flood forecast.摘 要 I Abstract II 目錄 IV 表目錄 VII 圖目錄 VIII 第一章 序論 1 1.1前言 1 1.2論文架構 3 第二章 文獻回顧 5 2.1類神經網路發展回顧 5 2.2自組織映射圖類神經網路相關文獻回顧 6 2.3類神經網路應用於水文研究之相關回顧 6 第三章 理論概述 8 3.1類神經網路概述 8 3.2類神經網路分類 9 3.2.1監督式與非監督式學習 9 3.2.2競爭式學習與非競爭式學習 11 3.3自組織映射圖類神經網路 12 3.3.1自組織映射圖 12 3.3.2迴歸輸出 16 3.4倒傳遞類神經網路 19 3.5 併聯式類神經網路架構 24 第四章 研究案例 29 4.1研究區域 29 4.2研究方法與評比指標 30 4.3模式架構 32 4.3.1自組織映射圖網路架構 32 4.3.2共軛梯度倒傳遞類神經網路架構 33 4.4輸入資料向量 34 4.5水文事件分析與併聯網路評比基準 36 4.6研究成果與討論 38 4.6.1下一時刻(t+1)蘭陽大橋流量預報模式 38 4.6.2下二時刻(t+2)蘭陽大橋流量預報模式 48 第五章 結論與建議 54 參考文獻 57 附錄A 64 附錄B 67 附錄C 88 附錄D 1012587314 bytesapplication/pdfen-US流量推估併聯式類神經網路倒傳遞類神經網路自組織映射圖CGBPSOMIncorporated Artificial Neural Networkthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/56115/1/ntu-93-R91622022-1.pdf