Enabling State Estimation for Fault Identification in Water Distribution Systems Under Large Disasters

We present a graphical model based approach for on-line state estimation of water distribution system failures during large-scale disasters. Water distribution systems often exhibit extreme fragilities during large-scale disasters (e.g., earthquakes) resulting in massive pipe breaks, water contamination, and disruption of service. To monitor and identify potential problems, hidden state information must be extracted from limited and noisy data environments. This requires estimating the operating states of the water system quickly and accurately. We model the water system as a factor graph, characterizing the non-linearity of fluid flow in a network that is dynamically altered by leaks, breaks and operations designed to minimize water loss. The approach considers a structured probabilistic framework which models complex interdependencies within a high-level network topology. The proposed two-phase approach, which begins with a network decomposition using articulation points followed by the distributed Gauss-Newton Belief Propagation (GN-BP) based inference, can deliver optimal estimates of the system state in near real-time. The approach is evaluated in canonical and real-world water systems under different levels of physical and cyber disruptions, using the Water Network Tool for Resilience (WNTR) recently developed by Sandia National Lab and Environmental Protection Agency (EPA). Our results demonstrate that the proposed GN-BP approach can yield an accurate estimation of system states (mean square error 0.02) in a relatively fast manner (within 1s). The two-phase mechanism enables the scalability of state estimation and provides a robust assessment of performance of large-scale water systems in terms of computational complexity and accuracy. A case study on the identification of "faulty zones" shows that 80% broken pipelines and 99% loss-of-service to end-users can be localized.