- 學位論文
以類神經網路預測自來水場混凝加藥量及其影響因子之研究
Investigation on the Effect Factors in Artificial Neural Networks for Predicting the Coagulant Dosage in Drinking Water Treatment Plants
摘要
類神經網路已成功應用在自來水場混凝加藥量之預測，然而如何提升預測能力，正是本研究探討的重點。本研究針對模糊理論、自身因子、資料正規化、時間區間對於類神經網路預測自來水場混凝加藥量之影響進行分析探討。為了解模糊理論之影響，本研究採用適應性模糊類神經網路與類神經網路進行比較有無模糊理論之影響。自身因子為預測目標值之前幾個單位時間的實測值。當使用的轉換函數不同時，則會影響資料正規化的需要性。 本研究資料分別來自台北縣某淨水場和台北市某淨水場。在可得到水質資料時，比較適合採用類神經網路模式來建立即時最佳混凝加藥量預測模式。在無法獲得進流水水質之情況下，則必須使用自身預測模式來預測即時最佳混凝加藥量，適應性模糊類神經網路模式較類神經網路模式佳。自身因子可提升類神經網路的預測能力，但增加自身因子的數目對於提升預測能力並不顯著。當隱藏層的轉換函數為雙彎曲線正切函數，輸出層的轉換函數為線性函數時，則輸入資料沒有正規化可使類神經網路得到較好的預測能力。短時間間距輸入資料可以提升類神經網路的預測能力。 當大雨導致高濁度時，類神經網路提供了操作者可以快速的改變最佳混凝加藥量，其模式輸入變數可透過Pearson相關係數分析來篩選，而在隱藏層的轉換函數為雙彎曲線正切函數，輸出層的轉換函數為線性函數，輸入資料不需要正規化。其輸入變數在台北縣某淨水場為原水濁度和自身因子(PAC(t-1))，在台北市某淨水場為原水濁度、分水井濁度、沉澱池濁度以及自身因子(D(t-1), D(t-2), D(t-3), 和 D(t-4))。
並列摘要
In the past, artificial neural network (ANN) has been used successfully for the prediction of coagulant dosage. Recent developments of ANN with the addition of fuzzy theory have also been reported. In this study, research focus is placed on how to improve the predictability of ANN, considering the effect of fuzzy theory, inherent factor, data normalization, and time interval on predicting the coagulant dosage in drinking water treatment. Performance of ANN and Adaptive Network based Fuzzy Inference System (ANFIS) approaches is compared in order to understand the effect of fuzzy theory in ANN. Inherent factor is defined as the past time object value. The use of different transfer functions determines the necessity of data normalization in ANN. Experimental coagulant dosage data are collected from the drinking water treatment stations in Taipei County and Taipei City, Taiwan. With raw water quality data made available in the analysis, ANN is suitable for building the real-time optimal coagulant dosage. On the other hand, the inherent predicting approach is useful to decide the real time optimal coagulant dosage, and the predictability of ANFIS is better than ANN. The inherent factor can improve the predictability of ANN, but increasing the amount of inherent factor is not significant for increasing the predictability of ANN. The predictability of ANN can be improved by 1) using a hyperbolic-tangent transfer function in the hidden layer, 2) using a linear transfer function in the output layer, 3) using the input data without data normalization, and 4) using short time interval input data. When a heavy rain let the raw water have high turbidity, ANN provides operators to decide the optimal coagulant dosage immediately. The input variables of ANN can be selected by the Pearson correlation coefficient, and the transfer function in the hidden layer is a hyperbolic-tangent function, and the transfer function in the output layer is a linear function, and input data is without data normalization. The input variables of drinking water treatment in Taipei County are raw water turbidity and inherent factor (PAC(t-1)), and the input variables of drinking water treatment in Taipei City are raw water turbidity, distribution water turbidity, sedimentation water turbidity, and inherent factor ((D(t-1), D(t-2), D(t-3), D(t-4)).