應用類神經網路預測科學園區污水廠二級出流水水質

Using Propagation Neural Network to Predict Secondary Effluent Quality from Science Park Sewage Treatment Plant

吳彬榮(Pin-lung Wu)

指導教授 ： 白子易

朝陽科技大學/理工學院/環境工程與管理系在職專班/碩士(2011年)

摘要

本研究採用倒傳遞類神經網路 (Back-Propagation Neural Network, BNN)，探討科學園區污水處理廠之出流水的水質變化。建立科學園區污水處理廠模擬架構，並對其出流水水質進行模擬預測，最後對模型預測值與實際值進行相對殘差之評析，以評估模型之預測效能。
在本研究之輸入層採取12個輸入變數，分別為進流水之酸鹼度 (potential of Hydrogen ,　pH)、污泥停留時間 (sludge retention times , SRT)、食微比 (Food to Microorganism ratio , F/M)、容積負荷檢核、固體物負荷、曝氣池停留時間、快混池停留時間、慢混池停留時間、沉澱池表面溢流率、沉澱池出流溫度、污泥容積指數（Sludge Volume Index , SVI）、懸浮固體濃度（Mixed Liquor Suspended Solids， MLSS)。輸出層採3個輸出變數，分別為出流水之懸浮固體（Suspended Solids,　SS）、生化需氧量(Biochemical oxygen demand , BOD)、化學需氧量 (Chemical Oxygen Demand , COD)進行個別預測。
本研究之操作速率為0.1、隱藏層神經元數分別為15個、訓練次數為10000次之條件下進行。在BOD部分於訓練過程中採取5個輸入變數對1個輸出變數時，出流水之平均絕對百分比誤差值（Mean Absolute Percentage Error, MAPE）為28.05%，相關係數 （Correlation Coefficient, R）為0.351；預測過程中則為採取12個輸入變數對1個輸出變數時，其MAPE值為23.06%、R值為0.269效果最好；在COD部分於訓練過程中採取12個輸入變數對1個輸出變數時，其MAPE值為19.49%、R值為0.420效果最好；預測過程中則為採取2個輸入變數對1個輸出變數時，其MAPE值為15.19%、R值為0.138效果最好；在SS部分於訓練過程中採取4個輸入變數對1個輸出變數時，其MAPE值為22.00%、R值為0.734效果最好；預測過程中則為採取4個輸入變數對1個輸出變數時，其MAPE值為20.23%、R值為0.168效果最好；在訓練過程中採取12個輸入變數對1個輸出變數時，出流水 SS、BOD、COD 之 MAPE 分別為22.63%、23.06%、19.49%， R 值分別為0.725、0.269、0.420。在預測過程中採取12個輸入變數對1個輸出變數時，出流水 SS、BOD、COD之 MAPE 分別為21.91%、23.06%、19.81%， R 值分別為-0.127、0.269、-0.025。整體而言，以BNN預測科學園區污水廠之SS、BOD及COD結果顯示MAPE值大部分都為合理範圍，表示對於預測濃度及變動趨勢皆可掌握，故利用類神經網路來預測工業園區污水廠出流水濃度是可行的。本研究相關結果可供科學園區污水處理廠操作診斷之參考。

關鍵字

倒傳遞類神經網路 ； 中部科學工業園區污水處理廠 ； 平均絕對百分比誤差值

並列摘要

This study used Back-Propagation Neural Network (BNN) to discuss quality changes of effluent water from science park sewage treatment plant in order to construct a science park sewage treatment plant simulation framework, and further simulate and forecast the effluent water quality. It finally evaluated the relative residual on the model predicted value and actual value to assess prediction efficiency.
This study adopted 12 input variables in the input layer, namely Potential of Hydrogen (pH) of entering water, Sludge Retention Times (SRT), Food to Microorganism ratio (F/M), Volume Load Review (VLR), Solids Load (SL), Aeration Basin Retention Times (ABRT), Fast Mixing Basin Retention Times (FMBRT), Slow Mixing Basin Retention Times (SMBRT), Sedimentation Basin Surface Overflow Rate (SBSOR), Sedimentation Basin Effluent Temperature(SBET), Sludge Volume Index (SVI), Mixed Liquor Suspended Solids (MLSS). Three output variables were adopted in output layer, which are Suspended Solids (SS) of effluent water, Biochemical Oxygen Demand (BOD), and Chemical Oxygen Demand (COD) to carry out individual forecast.
In this study, the operating rate was 0.1; the number of hidden layer neurons was 15, and the training frequency was 100,000. During the training course of BOD, when 5 input variables were adopted for 1 output variable, Mean Absolute Percentage Error (MAPE) of effluent water was 28.05%, Correlation Coefficient (R) was 0.351. During the forecast process, if 12 input variables were adopted for 1 output variable, the MAPE value was 23.06%, and the R value was 0.269, it has the best effect. During the training course of COD, if 12 input variables were adopted for 1 output variable, the MAPE value was 19.49%, and R value was 0.420, it has the best effect. During the forecast process, if 2 input variables were adopted for 1 output variable, the MAPE value was 15.19% and R value was 0.138, it has the best effect. During the training process of SS, if 4 input variables were adopted for 1 output variable, the MAPE value was 22.00%, and R value was 0.734, it has the best effect. During the forecast process, if 4 input variables were adopted to 1 output variable, the MAPE value was 20.23%, and R value was 0.168, it has the best effect. During the training course, if 12 variables were adopted to 1 output variable was adopted, the MAPE values of effluent water SS, BOD, COD were 22.63%, 23.06%, and 19.49%, respectively; R values were 0.725, 0.269, and 0.420, respectively. During the forecast process, if 12 input variables were adopted for 1 output variable, the MAPE values of effluent water SS, BOD, COD were 21.91%, 23.06%, and 19.81%, respectively; R values were -0.127, 0.269, and -0.025, respectively.As a whole, the results from forecasting SS, BOD and COD of science park sewage plant by BNN showed that most of MAPE values are in reasonable range, indicating that forecasting concentration and alteration trend may be both known. As a result, it is feasible to use BNN to forecast effluent water concentration of industrial park sewage plant. The relevant results of this study can be used as reference for operation and diagnosis of science park sewage treatment plant.

並列關鍵字

Back-Propagation Neural Network (BNN) ； Midland Science Industrial Park Sewage Treatment ； Mean Absolute Percentage Error (MAPE)

參考文獻

3. 阮成隆，「以倒傳遞類神經網路及多元線性迴歸模擬建築工地對台中縣粒狀污染物之影響」，碩士論文，朝陽科技大學，台中（2010）。

13. 蘇志朋，「以倒傳遞類神經網路及多元線性回規模擬建築工地對台中市粒狀污染物之影響」，碩士論文，朝陽科技大學，台中（2010）。

15. Pai T.Y., Y.P. Tsai, H.M. Lo, C.H. Tsai and C.Y. Lin. ”Grey and neural network prediction of suspended solids and chemical oxygen demand in hospital wastewater treatment plant Effluent”, Computers & Chemical Engineering. (2007a).

16. Pai T.Y., T.J. Wan, S.T. Hsu, T.C. Chang, Y.P. Tsai, C.Y. Lin, H.C. Su and L.F. Yu, “Using fuzzy inference system to improve neural network for predicting hospital wastewater treatment plant effluent,” Computers & Chemical Engineering. (2009a).

17. Pai T.Y., S.C. Wang, C.F. Chiang, H.C. Su, L.F. Yu, P.J. Sung, C.Y. Lin and H.C. Hu , “Improving neural network prediction of effluent from biological wastewater treatment plant of industrial park using fuzzy learning approach,” Bioprocess and Biosystems Engineering. (2009b).