Abstract
Wastewater treatment ensures that effluent water may safely be disposed of in bodies of water. During extreme weather conditions, the influent characteristics (ICs) stray from their typical behavior, affecting the performance of wastewater treatment plants (WWTPs). These points, referred to as anomalies, are discarded when modelling the WWTP performance. Considering this, an effluent quality prediction model, for a chlorination WWTP with moving bed biofilm reactor systems as secondary treatment, was developed. The ICs—flow rate, biological oxygen demand (BOD), chemical oxygen demand (COD), total coliform (TC), pH, and total suspended solids—were analyzed to determine the Mahalanobis distance (MD) and identify anomalies. The classification of MD was used to develop a support vector machine (SVM) model. Artificial Neural Networks (ANNs) were developed for both anomaly and non-anomaly points to predict the pre-chlorination and effluent BOD, COD, and TC. The optimal SVM for anomaly detection, modelled using 266 datapoints, 201 non-anomalies and 65 anomalies, used a fine Gaussian SVM architecture with an area under the receiver operating characteristic curve (AUC-ROC) of 0.90. The developed optimal ANNs exhibited correlation values ranging from 0.863 to 0.972. The generalization ability of the integrated SVM-ANN model was evaluated using a new set of data from the same WWTP. The mean absolute error values for the effluent BOD, COD, and TC value prediction were above average with values of 4.95 ppm, 12.18 ppm, and 26.29 MPN/100 mL. The overall model captures the trend of the test datasets, allowing the accurate forecast of effluent parameters and informing future modifications on WWTP design.
