應用倒傳遞與反傳遞類神經網路模式於洪流量之預測

Application of Back-Propagation Neural Network and Counter-Propagation Neural Network on the Flood Forecasting

陳昶憲(CHANG-SHIAN CHEN) ； 黃尹龍(YIN-LONG HUANG) ； 吳青俊(CHING-JIN WU) ； 蔡曜隆(YAO-LON TSAI)

《台灣水利》 49卷3期 (2001/09) Pp. 65-77

摘要

本文嘗試以類神經網路中之倒傳遞及反傳遞網路建立烏溪流域下一小時（一階）、二小時（二階）、三小時（三階）之洪流演算預報模式，其中將所蒐集的烏溪流域實測資料進行網路學習訓練，藉以推求模式代表性權重參數值。文中並嘗試分別以兩種不同的神經網路為基本架構，分別依輸入輸出處理單元數不同而建構單輸出與多輸出預測模式。文中所建立之預測模式皆嘗試於應用於烏溪流域。經預測準確度驗證，其評鑑結果若以精度要求或工程實際應用上皆有其可取之優點。文中並另比較分析BPN與CPN在同一網路架構所採取的單輸出模式與多輸出模式兩種預測模式的預測精確度及優缺點，以提供後人使用之參考。

關鍵字

線性轉換函數（LTF） ； 倒傳遞網路（BPN） ； 反傳遞網路（CPN）

並列摘要

In this paper, Back Propagation Network and Counter Propagation Network have been employed to establish the one to three step ahead forecasting. The data collected from Wu-Shi Basin have been used to proceed the training of leaning network and obtain the model parameters thereby. In this study, we tried two kinds of different neural network for the basic structure, according to the difference input outputs to treat unit different while we can build several forecasting models. Further, the recalling process of flood forecasting model was often performed in order to forecast the one to three step ahead outflow of downstream gauging. The results obtained model verification are satisfactory. Comparisons on the simulation results were also made for these two models in this study.

並列關鍵字

Linear transfer function (LTF) ； Back-propagation neural network (BPN) ； Counter-propagation network (CPN)