Development of prediction model for chlorination wastewater treatment plants using artificial neural networks with Mahalanobis distance - based support vector machine

• Main Author: Jaluague, Andrei Fryle
• Format: text
• Language: English
• Published: Animo Repository 2022
• Subjects: Sewage disposal plants; Sewage—Purification—Chlorination; Chlorination; Chemical Engineering; Engineering
• Online Access: https://animorepository.dlsu.edu.ph/etdb_chemeng/24; https://animorepository.dlsu.edu.ph/context/etdb_chemeng/article/1022/viewcontent/2022_Jaluague_CompleteVersionETD.pdf
• Institution: De La Salle University

The performance of chlorination wastewater treatment plants (WWTPs) must be determined to identify its effectiveness in reducing pollutants in wastewater. This is directly affected by influent conditions (ICs), which reflects the behavior of the plant’s external environment. These effects were integrated with artificial neural network (ANN) modeling through Mahalanobis distance-based support vector machine (SVM), an anomaly detection algorithm. The analysis was done for two chlorination WWTPs: one with a moving bed biofilm reactor system (MBBR), while the other with conventional activated sludge (CAS) system for secondary treatment. Optimal SVM networks for classifying anomalies and non-anomalies were created using 266 and 221 datapoints, for the MBBR and CAS systems, respectively. Both utilized a fine Gaussian SVM architecture resulting in an area under the receiver operating characteristic curve of 0.90 (MBBR system) and 0.86 (CAS system). ANN networks were created to predict values for effluent biological oxygen demand (BOD), chemical oxygen demand (COD), and total coliform (TC), based on their SVM classification. The optimal networks had different transfer functions and network architectures reinforced the importance of specialized networks for different ICs. The performance of all ANNs was identified through their correlation coefficient (R), ranging from 0.818 to 0.997. The generalization capabilities for pairing ANN with SVM was evaluated using a new set of data from the same WWTP cases. The mean absolute error values for MBBR and CAS, respectively, were 4.95 and 2.32 ppm for BOD, 12.18 and 18.84 ppm for COD, and 26.29 and 198.96 MPN/100 mL for TC. The framework was able to capture the trend of the test datasets, reinforcing its ability for effluent parameter prediction for both case studies. Its errors were attributed to the presence of overfitting, lack of datapoints representing anomaly ICs, an infrequent grab sampling method, and exclusion of process parameters.