應用深度學習預測污水處理廠曝氣槽之溶氧

Prediction of Dissolved Oxygen in Aeration Tank of Wastewater Treatment Plant Based on Deep Learning Algorithm

鄭婷方(Ting-Fang Jheng)

指導教授 ： 駱尚廉

國立臺灣大學/工學院/環境工程學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202000043

摘要

在污水處理廠中，維持設備的正常運轉需要消耗大量的電力，而二級處理中的曝氣槽更是位居消耗電能首位，佔所有電力成本的45-75%。透過對曝氣槽中可即時監測之數據進行溶氧預測，可達到有效控制曝氣槽電力的消耗，並且能夠增加曝氣槽針對水質淨化的效率。因此，本研究將利用以往及目前較為熱門之模型針對曝氣槽溶氧預測效能進行比對，並且選出該曝氣槽最合適的預測模型。
本研究使用桃園龜山污水處理廠2018年之每日實驗數據取pH值、溫度、COD與SS為輸入之特徵值進行模型的建立，分別使用多元線性迴歸，XGBoost、深度神經網絡進行溶氧分類預測。由於資料庫的數量不足，本研究於前處理階段使用高斯白噪音擴增數據至每筆特徵值2184筆。而後使用多元線性迴歸套入一次項與二次項的特徵值分類預測，結果顯示一次項之模型得到最好的預測準確率為85.81%，而AUC (Area under ROC curve) 為77.95%。再者，利用XGBoost建立模型所獲得之準確率達92.45%，AUC為87.61%。最後，在利用深度神經網絡兩層隱藏層與三層隱藏層中，可得到最佳化的模型為兩層隱藏層之DNN，第一層隱藏層為95個節點，第二層為115個節點，此模型在訓練與驗證準確率分別達到86.83%與82.15%，AUC得到78.83%。由三個模型比較，可發現利用XGBoost所建立之模型可較有效的預測溶氧的分類；此外，透過非線性模型之建立，更能捕捉二級處理中曝氣槽的水質特性。

關鍵字

高斯白噪音 ； 深度學習 ； 多元線性迴歸 ； XGBoost ； 深度神經網絡

並列摘要

In employing wastewater treatment, maintaining equipment functionality requires a large amount of electricity. Furthermore, the aeration tank in the secondary treatment consumes the highest portion, accounting for 45-75% of all electricity costs. Through the prediction of dissolved oxygen (DO) by the data that can be real-time monitored in the aeration tank, the consumption of the aeration tank power can be effectively controlled. Moreover, the efficiency of the aeration tank for water purification can be improved. Therefore, the objective of this study is to compare effectiveness of popular models from the past and present in predicting DO in the aeration tank in terms of classification, and finally select the most suitable model for the aeration tank.
In this study, daily experimental data of the wastewater treatment plant in Taoyuan, Taiwan during 2018 is used to establish the models. Values of pH, temperature, COD and SS are considered as related parameters to DO and are used as model inputs. In this study, three models are utilized for comparison: multiple linear regression, XGBoost and deep neural network (DNN). At first, the database is extended from 365 sets to 2184 sets by adding Gaussian white noise to the measurements. Then, multiple linear regression is used to predict the classification of DO based on different polynomial functions. The result shows that using model with 2nd order polynomial function predicts best performance with accuracy of 85.58%, and AUC (Area under ROC curve) of 76.52%. In addition, the accuracy of using XGBoost is 89%, and with AUC up to 87.30%. Finally, DNN with two hidden layers and three hidden layers are constructed and compared, the optimized model can be obtained when DNN model consists of two hidden layers. The results show that the accuracy of training and validation of the model reaches 86.83% and 82.15%, respectively, while the AUC yielded 78.83%.
As a result, the model established by XGBoost can more effectively predict the classification of DO than the other two types of models. Moreover, the water quality characteristics of the aeration tank in the secondary treatment can be better extracted using a nonlinear model.

並列關鍵字

Gaussian White Noise ； Deep Learning ； Multiple Linear Regression ； XGBoost ； Deep Neural Network ； Dissolved Oxygen Prediction

參考文獻

Abyaneh, H. Z. (2014). Evaluation of multivariate linear regression and artificial neural networks in prediction of water quality parameters. Journal of Environmental Health Science and Engineering, 12(1), 40.

Adler, R. W., Landman, J. C., and Cameron, D. M. (1993). The clean water act 20 years later. Island Press.

Çamdevýren, H., Demýr, N., Kanik, A., and Keskýn, S. (2005). Use of principal component scores in multiple linear regression models for prediction of Chlorophyll-a in reservoirs. Ecological modelling, 181(4), 581-589.

Chau, K. W., Cheng, C. T. (2002). Real-time prediction of water stage with artificial neural network approach. Lecture Notes in Artificial Intelligence 2557, 715.

Chau, K. W. (2006). A review on integration of artificial intelligence into water quality modelling. Marine pollution bulletin, 52(7), 726-733.