A Machine Learning Framework for Overflow Prediction in Combined Sewer Systems

Abstract

Sewer overflows, which often occur during heavy rainfall events, threaten public health and the environment, contributing to drinking water contamination and other concerns. Knowing ahead of time the location and volume of overflows is critical for sewer operators to make decisions to mitigate these overflows. Accurate and ultrafast models to predict the location and volume of sewer overflows are essential for optimal sewer operation (e.g., by load balancing). Machine learning (ML) is a rising field with an excellent capacity for predicting outcomes in different applications. This paper proposes a new methodology that bridges big data algorithms and physics-based numerical models to improve sewer overflow predictions. The combined sewer system (CSS) of the Puritan-Fenkell 7-mi facility in Detroit, MI, is chosen as the prototype in this study. Our model used deep learning (DL) approach for time-series forecasting. The rainfall-overflows relationship in this study is predicted by various deep learning models. This data-driven model is trained on various return period data sets and tested by an unused return period data set. Overall, our ML framework can accurately predict the location and time traces of sewer overflow volumes in a time frame of seconds.