Machine Learning–Based Source Identification in Sewer Networks

Abstract

Motivated by the valuable epidemiological information it reveals, wastewater surveillance has received significant attention in recent years. Furthermore, monitoring the water quality in sewer systems has been shown to provide useful information to support wastewater treatment operations. Yet, a critical need still exists for developing novel approaches for rapid and efficient source identification of chemical and biological species of interest in sewer systems. A limited number of source identification approaches have been proposed in previous literature, and the majority of these approaches employed various simplifying assumptions that limit their usage in real-life applications. In this study, a machine learning–based simulation-optimization framework was developed to determine the characteristics (i.e., concentration and loading pattern) of multiple simultaneous injection sources in sewer systems. The simulation was conducted using a surrogate model in the form of a multilayer perceptron neural network, which was trained using simulation results derived from the Storm Water Management Model (SWMM). The simulation model was then coupled with a genetic algorithm to reveal the characteristics of multiple sources that reproduce the concentration patterns observed at one or more monitoring locations in the sewer system. The proposed framework was applied to a range of injection scenarios and was able to identify the characteristics of multiple simultaneous injection sources under different conditions. The results showed that the residence time plays a significant role in the identifiability of the injection source location. The proposed framework is applicable to a wide number of source identification applications, including contamination source identification and wastewater-based epidemiology.