坡地集水區類神經網絡模式之研究及其應用於水文之預測

本文之目的乃應用類神經網絡之概念，以建立合適之水文模式，用以從事本省坡地集水區內降雨一逕流歷程之模擬。一般坡地集水區系統常呈地形嶇崎、流路繁複，有如神經網絡之特性，故乃應用類神經網絡理論中倒傳遞演算法（卽BP演算法）之學習能力與時間序列之觀念以建立一種具有輸入、隱藏及輸出等三層之時間序列類神經網絡模式，來模擬並預測非線性之水文現象。該模式啓始於學習輸入（流量及雨量）與輸出（流量）資料間之關係，並求得其間之特性係數；經完成學習之程序後，僅需輸入資料，便可預測輸出值。在模式架構方面，文中分別以單輸出與多輸出來建模，並套配不同之隱藏層單元數目計算方法所得之A、B及C型三種不同組合，以模擬颱洪事件之逕流歷程。本文以德基水庫上游集水區之颱洪事件作爲模擬之依據，並得到令人滿意之驗證結果。因此，本模式可進一步應用於上游集水區颱洪事件之模擬，俾供臺灣集水區防洪規劃及水土保持治理之參考應用。

關鍵字

類神經網絡；時間序列；倒傳遞演算法；網絡層單元數目

並列摘要

The main purpose of this paper is to set up an adequately hydrological procedure for simulating the rainfall-runoff process on the slopeland watersheds in Taiwan by applying the model of artificial neural network, abbreviated as ANN. On the basis of learning capability in ANN and the conception of time series, a three-layer time series artificial neural learning network model with input, hidden and output layers is devised to forecast runoffs and formulate the non-linear hydrological phenomena. The back-propagation algorithm is introduced to conduct the training process to recognize the correspondence between inputs and outputs. After the training process is performed, the model proposed in this study could be applied to forecast runoffs with only the input data and calibrated model weights. Multiple input-single output approach and multiple input-multiple output approach, in which the output units are 3,4 and 5 respectively, are regarded as two different categories of the model proposed to simulate the rainfall-runoff process during the flooding period. To verify the appropriateness of the two approaches adopted, the slopeland watershed of the Te-Chi Reservoir in middle Taiwan is chosen as a project area. The results of the two approaches proposed show a satisfactory simulation. Therefore, it is justifiable to command that this study can be further employed to estimate and forecast the peak of flood during the typhoon period in slopeland watersheds and play an important role on the planning of flood mitigation in Taiwan.

並列關鍵字

Artificial neural network ； Time series ； Back-propagation algorithm ； Number of hidden nodes

主題瀏覽