Abstract
Stormwater networks are critical utility infrastructures designed to drain rainwater and nuisance flows, such as excess irrigation and groundwater seepage from urban communities. During this process, they can transport pollutants (e.g., pesticides, oils, and greases) to receiving waters such as rivers, bays and oceans. A recurring problem faced by these systems are dry weather flows (DWFs), where illicit discharges are introduced and propagated in the network during periods with no rain. Current techniques for monitoring DWFs consist of manual inspections and grab samples, which are costly and inefficient. However, with advances in sensing and communication, the Internet-of-Things (IoT) has enabled new opportunities for enhanced decision support and control. This paper proposes a quick and efficient physics-based backwards inference model to identify potential sources of pollutant discharges in DWFs, given time-series IoT observations and knowledge embedded in domain-expert simulations. Our approach leverages the underlying physics that drives flow propagation in stormwater systems, and optimizes multiple least-squares regressions to find potential DWF sources and their associated flows. We evaluate our backwards inference model on six real-world stormwater networks provided by domain experts, and show its efficacy in reconstructing anomalies.
