查詢結果分析
來源資料
相關文獻
- 應用於洪水演算類神經倒傳遞網路法最適參數推估
- Nonlinear Prediction of Daily Streamflow Time Series Models in the Hualien Area
- 應用霍普菲爾神經網路解考慮污染量之經濟調度問題
- Innovative Power System Damping Control Strategies for the Unified Power Flow Controller (UPFC) Based on Fuzzy Control Theory
- Vector Quantization in DCT Domain Using a Grey Competitive Hopfield Neural Network
- 以類神經網路預測河川短期距流量
- 應用模糊控制進行洪流量預報
- 巢狀超矩形學習模式於乾旱季節流量預測之應用
- 以類神經網路預測玉峰橋站高水位時期之流量
- 集水區旬流量預測研究
頁籤選單縮合
題 名 | 應用於洪水演算類神經倒傳遞網路法最適參數推估=Estimating the Optimal Parameters of a Back-Propagation Network on Flood Forecasting |
---|---|
作 者 | 陳昶憲; 蔡國慶; | 書刊名 | 逢甲學報 |
卷 期 | 33 1998.06[民87.06] |
頁 次 | 頁39-56 |
分類號 | 443.42 |
關鍵詞 | 倒傳遞類神經網路法; 學習速率; 慣性因子; 效率係數; 流量預測; 能量函數; Back-propagation network; Leaning rate; Momentum factor; Coefficient of efficiency; Flood forecasting; Energy function; |
語 文 | 中文(Chinese) |
中文摘要 | 本文有鑑於在使用類神經網路法進行流量預測時,由於水文參數一般皆具逢機特 性,因而有學習(檢定)結果良好,但是預測結果卻不佳之情形,為了令預測模式具實用性 ,因而進行參數的敏感度分析,以找尋較合適之學習速率、慣性因子、學習效率係數值,以 使模式達到收斂快預測準的目的。本文以烏溪流域之水文測站為實例演算,進行流量預測之 比較,發現進行學習演算時,學習之精度不需要求太高,學習效率係數到達 0.92, 即可得 到較佳之預測結果,學習精度要求過高,反而會造成過度學習,而使加權值學習強迫去適應 學習範例之曲線。而學習速率與慣性因子之搭配使用上,建議使用較大的學習速率搭配較小 的慣性因子,例如:η =1.0 ∼ 0.8、α =0.2 或更小,將會有較好的預測結果。 |
英文摘要 | In this study, a flood forecasting model was established by a back- propagation neural network. Because of the random characteristics of hydrological parameters, a good model learning (calibration) case may follow with a poor recalling (verification) process. In order to improve this defect and increase the practicability of the model, the sensitivity analysis was proposed to obtain the optimal values of the learning rate, momentum factor, and efficiency of learning such that the model can converge quickly and forecast precisely. In the case study of Wu-Shi basin in which the comparison of discharge forecasting was made, an accurate forecasting result was obtained without high quantity of learning. A moderate number, 0.92, was used for learning rate. If the accuracy of learning is highly demanded, over learning may occur and weights will be forced to only suit the learning case. Additionally, in order to have better forecasting results, if the learning rate is increased up to unit, the momentum factor should be decreased to 0.2 or the smaller. |
本系統中英文摘要資訊取自各篇刊載內容。