Coupling machine learning and agent-based modeling to characterize contamination sources in water distribution systems for changing demand regimes

Water distribution systems (WDSs) deliver clean, safe drinking water to consumers, providing an essential service to constituents. WDSs are increasingly at risk of contamination due to ageing infrastructure and intentional acts that may be possible through cyber-physical vulnerabilities. Identifying the source of a contamination event is challenging due to limited system-wide water quality monitoring and non-uniqueness present in solving inverse problems to identify source characteristics. In addition, changes in the expected demand patterns that are caused by, for example, social distancing during a pandemic, adoption of water conservation behaviors, or use of decentralized water sources, can change the anticipated propagation of contaminant plumes in a network. This research develops a computational framework to characterize contamination sources using machine learning (ML) techniques and simulate water demands and human exposure to a contaminant using agent-based modeling (ABM). An ABM framework is developed to simulate demand changes during the COVID-19 pandemic. The ABM simulates population movement dynamics, transmission of COVID-19 within a community, decisions to social distance, and changes in demands that occur due to social distancing decisions. The ABM is coupled with a hydraulic simulation model, which calculates flows in the network to simulate the movement of a contaminant plume in the network for several contamination event scenarios. ML algorithms are applied to determine the location of source nodes. Research results demonstrate that ML using random forests can identify source nodes based on in-line and mobile sensor data. Sensitivity analysis is conducted to explore the number of mobile sensors that are needed to accurately identify the source node. Rapidly identifying contamination source nodes can increase the speed of response to a contamination event, reducing the impact to the community and increasing the resiliency of WDSs during periods of changing demands.