中 文 摘 要
台灣地理位置位處氣象之亞熱帶，屬於多雨地區，由於降於台灣之雨量在空間與時間上分佈不均，加上台灣地勢陡峻，河川流短，地質薄弱，導致每次一到颱風季節，颱風帶來的豪雨常氾濫成災！因此，如何防洪治水降低災害，是為當務之急。
以往對洪水的防治著重於工程方法，時至今日，防洪治水不只注重工程建設，更加重視非工程層面的早期預報及預警，冀能預估洪泛水位而提早應變，將洪災造成之損失減至最低。常用的水位預報及推求模式大多屬定率模式，模式中所包含的水文變量及不確定因子如河川流量、水位、坡度、濕周、河道的斷面積、底床糙率、渦流損失係數、動量修正係數、曼寧係數、蔡斯係數…等，在求解時需做許多假設，輸入因子若有錯估，其預測結果與實際值將有相當誤差，有鑑於此，本文提出以GMDH (Group Method of Data Handling) 理論建立一單純的水位輸入~輸出模式，替代傳統複雜之水文演算來預報河川水位。如配合現場水位站資料自動化蒐集與傳輸系統擷取更新的水位數據作即時輸入，即可建立遞迴的GMDH修正模式，使預測模式具時變性而能自我調整，達到長期觀測、精確預測的效果。
本文以曾文溪流域之玉田水位站、麻善大橋水位站、將軍潮位站及曾文新村雨量站蒐集之數據建模作水位預測之實例研究，首先以試誤法找出最佳建模筆數，並以逐步迴歸之GMDH (SGMDH) 來改善原GMDH演算法會衍生高階非線性項而降低其實用性的缺點。模式經檢定及驗證後，再利用該預測模式作賀伯颱風、701暴雨、807暴雨、608暴雨及809暴雨五場暴雨情況麻善大橋水位站前置時間1-6小時之水位預報，大體來說，前置時間4~5小時之水位預測在允許誤差10%之範圍內；本文同時以608暴雨建模去預測其它四場暴雨，所得之結果亦甚為滿意。本文又將預測結果與時間序列之ARIMA模式作一比較，由比較結果可知，本模式在低水位時之預測稍不如ARIMA模式，但本模式在洪峰水位的預測上，卻遠優於ARIMA模式，若從洪水預報著重高水位之觀點來看，本模式具有較好的預測效能。
本文最後提出即時水位預報系統之建置構想，利用本模式配合本系所自1988年即已發展並建置於現場之有線/無線數據自動化傳輸系統，即時取得現場資料並納入預報模式中演算分析，即可建立即時水位預報系統模式，在洪水來臨前能預報洪泛水位，事先提出預警，迅速提供相關數據予決策者擬定應變措施之參考。

ABSTRACT
Taiwan locates in subtropical zone of meteorological condition and gets plenty of raining in monsoon and typhoon season. Because of the unequal rainfall distribution in space and period and the steep slope of streambed in Taiwan environment, heavy storm usually cause serious suffering during monsoon / typhoon season. For example, the Herb typhoon that occurred in July 1996 caused the most disaster included almost all over Taiwan area since 1961. Hence, flood prediction, control and mitigation in order to prevent the suffering from flood and reduce the calamity is the imperative duty of flood control authorities.

Engineering construction is dominantly considered in early period to against floods. Establishing the flood forecasting and early warning system pay much attention by now to prevent and minimize the damage caused by flood. Hydrologic / flood routing were used usually to calculate and predict the water level in stream. Hydraulic parameters such as manning coefficient, eddy viscosity coefficient, side flow, water level, discharge, average velocity and cross section of specific stream have being used as the input to solve this water level forecasting problem. The uncertainty of these hydraulic parameters might cause prediction error or calculation divergence while these traditional models were in computational procedure. A framework based on GMDH (Group Method of Data Handling) is proposed in this paper to establish the I/O model as the alternative by using relatively simple field measuring water / tidal level and rainfall data as the model input and predict the prior hours water level of specific river during storm period. Then, SGMDH (stepwise regression GMDH) algorithm has been introduced in this model to improve the GMDH algorithm that can easily induced the high order nonlinear terms in calculating step so as to reduce the benefit of predicting procedure. The update water level and rainfall data were collected to organize the Sequential GMDH modified model to yield the time variant properties in stream surface level forecasting.

Water level records of Yu-Ten station as well as Ma-Shan station of Tseng-Wen Chi, tidal level record of Jiang-Jiun station and rainfall data of Tseng-Wen Village have been chosen as the main input to establish and updating. Five serious rainstorms (Herb typhoon, 701 rainstorm, 807 rainstorm, 608 rainstorm and 809 rainstorm) have been tested with this forecasting model to predict the water level of Ma-Shan station by using the prior 1 to 6 hours’ leading step of water / tidal level and rainfall data. Analysis results show that the forecasting errors were within 10% with the prior 4 to 5 hours’ leading step of data input. The GMDH forecasting model established by using the data of 608 rainstorm to predict the other 4 rainstorms obtained satisfied forecasting results.

An ARIMA (Auto Regressive Integrate Moving Average) model was provided as well to predict the water level of Ma-Shan station and compared with the results obtained from GMDH forecasting model in this study. The simulate results show that ARIMA model is little superior in low water level prediction and GMDH forecasting model get better forecasting results than ARIMA model in high water / flood level prediction.

Finally, an automatic data acquisition and transmission system developed by Field Survey Laboratory of Hydraulics and Ocean Engineering of National Cheng Kung University was promoted in-situ to employ the real time stream surface level, tidal level and rainfall data and enhanced the data input function for model calibration, verification and parameters modification. This real time GMDH forecasting model cooperate with this system as mentioned before provide the facilities of picking up field data and transmitting it to the flood control center as the boundary data input and updating this time variant forecasting model. Good prediction results could be obtained by this GMDH forecasting model through the verification of field data, hence the GMDH forecasting method proposed in this paper can get the excellent validity in water level prediction during heavy storm period.

參 考 文 獻
1.顏月珠，「現在統計學」，三民書局。
2.吳柏林，「時間數列分析導論」，華泰書局。
3.林真真、鄒幼函，「迴歸分析」，華泰書局。
4.吳賢銘、潘迪特，「時間數列與系統分析」，中央圖書出版社。
5.王如意、易任，「應用水文學」，新編上下冊，國立編譯館出版。
6.張斐章、黃源義、徐國麟，「自組性演算法及其應用於降雨-逕流模式建立之研究」，台大農學院研究報告，pp.293-316，1994。
7.張斐章、梁晉銘，「自組性演算法於河川水位流量預測之研究」，台灣水利，第41卷，第四期，pp.14-24，1993。
8.林淑真、李宗仰，「自我聯合模式於感潮河段之水位預報」，土木水利，第32卷，第四期，pp.3-16，1996。
9.曾財益，「洪水即時預報模式及不確定性分析之研究」，碩士論文，國立成功大學水利暨海洋工程研究所，1995。
10.顏銘佑，「現場調查資料擷取與傳輸自動化之研究」，碩士論文，國立成功大學水利暨海洋工程研究所，1994。
11.李梅芳，「自組非線性系統應用於水位相關之研究」，碩士論文，國立成功大學水利暨海洋工程研究所，1993。
12.顏沛華，「地下水位及地層下陷監測無線/有線數據自動化傳輸之研究」，工業技術研究院委託學術機構研究計劃-期末報告，1998。
13.羅積玉，「多元統計分析方法與應用」，科技圖書股份有限公司。
14.楊永祺，「自組非線性系統應用於水位即時預測」，碩士論文，國立成功大學水利暨海洋工程研究所，1997。
15.周益達，「自組非線性理論建立水位預測模式之探討」，碩士論文，國立成功大學水利暨海洋工程研究所，1998。
16.李文揚，「自組非線性系統應用於水庫優氧化之預測」，碩士論文，國立成功大學水利暨海洋工程研究所，2000。
17.吳俊逸，「自組非線性系統應用於地層下陷之預測」，碩士論文，國立成功大學水利暨海洋工程研究所，2001。
18.蔡耀賢，「消滅下游河道洪峰之水庫防洪操作分析」，碩士論文，國立成功大學水利暨海洋工程研究所，1998。
19.陳伶姮，「河川旬流量時間序列之研究—以石門水庫入流量為例」，國立海洋大學河海工程學系，2001。
20.翁俊鴻，「水文時間序列之自相關結構及其潛在應用」，碩士論文，國立成功大學水利暨海洋工程研究所，1994。
21.虞國興、莊明德，「長記憶水文時序模式之探討」，台灣水利，第37卷，第四期，pp.37-49，1989。
22.鄭子璉、周乃昉，「徐昇多邊形網法之數值計算」，台灣水利，第48卷，第三期，pp.43~51，2000。
23.蔡長泰，「曾文溪洪水預報模式之研究(一)」，行政院國家科學委員會防災科技研究報告80-80號，1993。
24.蔡長泰、蔡智恆，「洪流演算模式(FRLFI-模式)之建立(Ⅲ)使用者手冊」，行政院國家科學委員會專題研究計畫成果報告，1994。
25.謝世傑、李戎威、黃巨松，「淡水河流域洪水預警系統更新、改善與展望」，水利，第1卷，第十期，pp.35~46，2001。
26.黃慶光、張雲羽、鍾鴻霖、林杰熙、夏漢民、鄭昌奇、王瑞雯、楊淑敏，「洪水預報系統整合規劃」，第十二屆水利工程研討會，pp.D-25~D-31，2001。
27.A.G.Ivakhnenko，「Heuristic Self-Organization in Problem of Engineering Cybernetics」，Automatic，Vol.6，1970，pp.207-219。
28.Ikeda S.,S. Fugishige,and Y. Sawaragi，「Nonlinear Prediction Model of River Flow by Self-Organization Method」，Int. J. Syst. Sci.，Vol.7，No.2，pp.165-176，1976。
29.Peter Young、Paul Whitehead，「A recursive approach to time-series analysis for multi-variable systems」，INT.J.Control Vol.25，No3，pp.457-482，1977。
30.Saburo Ikeda、Satoru Fujishige and Yoshikazu Sawaragi，「Non-linear prediction model of river flow by self-organization method」，INT.J.Systems Sci.Vol.7，No.2，pp.165-176，1976。
31.Stanley J. Farlow，「Self-organizing Methods In Model-GMDH TYPE Algorithms」，MARCEL DEKKER.INC。
32.Yoshimura ,T.,R. Kiyozumi, K. Nishino ,and T. Soeda，「Prediction of Air Pollutant Concentrations by Revised GMDH Algorithms in Tokushima Prefecture」，IEEE Trans. Syst., Man ,Cybern.，Vol.SMC-12，No.1，pp.50-56，1982。
33.Duffy ,J.J. and M. A. Franklin，「A Learning Identification Algorithm and Its Application to an Environmental Systems」，IEEE Trans. System，Man，Cybernetic，Vol.SMC-5，No.2，pp.226-239，1975。
34.Tamura，H. and T. Kondo，「Heuristics Free Group Method of Data Handling Algorithm of Generating Optimal Partial Polynomials with Application to Air Pollution Prediction」，Int. J. Syst. Sci.，Vol.11，No.9，pp.1095-1111，1980。

網路(WWW)參考
1.經濟部水利署
http://www.wra.gov.tw/spring.asp
2.中央氣象局資訊服務網
http://www.cwb.gov.tw/v3.0/index.htm