AUTHOR=Thompson Kyle A. , Branch Amos , Nading Tyler , Dziura Thomas , Salazar-Benites Germano , Wilson Chris , Bott Charles , Salveson Andrew , Dickenson Eric R. V. TITLE=Detecting industrial discharges at an advanced water reuse facility using online instrumentation and supervised machine learning binary classification JOURNAL=Frontiers in Water VOLUME=4 YEAR=2022 URL=https://www.frontiersin.org/journals/water/articles/10.3389/frwa.2022.1014556 DOI=10.3389/frwa.2022.1014556 ISSN=2624-9375 ABSTRACT=

Industries occasionally discharge slugs of concentrated pollutants to municipal sewers. These industrial discharges can cause challenges at wastewater treatment plants (WWTPs) and reuse systems. For example, elevated total organic carbon that is refractory through biological wastewater treatment increases the required ozone dose, or even exceeds the capacity of the ozone unit, resulting in a treatment pause or diversion. So, alert systems are necessary for potable reuse. Machine learning has many advantages for alert systems compared to the status quo, fixed thresholds on single variables. In this study, industrial discharges were detected using supervised machine learning and hourly data from sensors within a WWTP and downstream advanced treatment facility for aquifer recharge. Thirty-five different types of machine learning models were screened based on how well they detected an industrial discharge using default tuning parameters. Six models were selected for in-depth evaluation based in their training set accuracy, testing set accuracy, or event sensitivity: Boosted Tree, Cost-Sensitive C5.0, Oblique Random Forest with Support Vector Machines, penalized logistic regression, Random Forest Rule-Based Model, and Support Vector Machines with Radial Basis Function Kernel. After optimizing the tuning parameters and variable selection, Boosted Tree had the highest testing set accuracy, 99.2%. Over the 5-day testing set, it had zero false positives and would have detected the industrial discharge in 1 h. However, setting fixed thresholds based on the maximum normal datapoint within the training set resulted in nearly as good testing set accuracy, 98.3%. Overall, this study was a successful desktop proof-of-concept for a machine learning-based alert system for potable reuse.