倒傳遞類神經網路於淨水混凝自動加藥前饋控制應用之研究-模廠試驗

Automatic Coagulant Dosing System by Back Propagation Artificial Neural Network (BPANN) in Water Treatment Plant Operation

陳大為(Ta-Wei Chen)

指導教授 ： 黃志彬

國立交通大學/工學院/工學院永續環境科技學程/碩士(2012年)

https://doi.org/10.6842/NCTU.2012.00039

摘要

混凝為傳統淨水場之主要處理單元，混凝加藥量多由操作人員透過瓶杯試驗及個人現場操作經驗決定，往往無法即時調整準確之混凝劑量，造成混沉效能不彰。本研究主要以實驗室瓶杯試驗建置不同水質參數與最適加藥量之數據庫，藉此建立倒傳遞類神經網路(back propagation artificial neural network, BPANN)預測混凝劑量之最佳模式，再以具傳統處理單元之模廠處理天然濁水及人工高濁水，進行BPANN前饋自動加藥控制系統效益驗證。
研究結果顯示，使用實驗室瓶杯試驗114筆數據作為BPANN之訓練、驗證及測試資料，在LM演算法、隱藏層1層及早停止學習原則參數組合下，以原水濁度、pH、溫度及鹼度構成不同輸入參數組合所建立之三組BPANN模式，其測試相關係數(r)均可達0.93以上。在模廠試驗中，當天然原水濁度在100 NTU以下或人工高濁水濁度在1,000 NTU左右時，測試三組建立之BPANN模式，僅有水質輸入參數為原水濁度、pH二項所建立之BPANN模式，可即時準確反應模廠原水水質所需之混凝劑量，處理水質亦符合實場內控出水水質標準。

關鍵字

水處理 ； 混凝 ； 類神經網路 ； 自動加藥

並列摘要

Coagulation is an essential unit in a conventional water treatment plant (WTP), of which the dosage is generally determined by jar tests and the experiences of the operators. Such an operation often results in inaccurate dosing and poor performance. In this study, back propagation artificial neural network (BPANN) was applied in the prediction of coagulant dosage. The best model of BPANN was first established from the data base containing various parameters of water quality and the corresponding optimum dosage generated from the lab-scale jar tests. The efficiency of the automatic feed-forward dosing system by BPANN was verified by a pilot-plant with conventional water treatment units targeting natural water and synthetic high turbidity water.
Results of 114 jar tests were used to train, validate and test the BPANN. When the combination of LM calculation and one hidden-layer were set in the principal of early stop, all the relative coefficients (r) of prediction for various BPANN patterns edited by inputting different combinations of turbidity, pH, temperature and alkalinity of raw water exceeded 0.93. In the pilot study for the coagulations of natural water of below 100 NTU and synthetic turbidity water of around 1000 NTU, only the BPANN automatic dosing system validated by two parameters, namely, turbidity and pH, made a real-time response to predict the correct dosage for coagulation and meet the water quality standard of the WTP.

並列關鍵字

water treatment ； coagulation ； artificial neural network ； automatic dosing

參考文獻

吳冠德，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，國立台灣大學環境工程學研究所博士論文，2009。

蔡瑞煌，類神經網路概論，三民書局，1995。

Bae, H., Kim, S. and Kim, Y.J., (2006) “Decision algorithm based on data mining for coagulant type and dosage in water treatment systems ” Water Science & Technology 53(4–5), 321–329.

Cheng, W.P., Kao, Y.P. and Yu, R.F. (2008) “A novel method for on-line evaluation of floc size in coagulation process” Water Res., 42, 2691–2697.

Dentel, S. K. and Kingery, K. M.. (1989) “Using streaming current detectors in water treatment” J. AWWA, 85–94