Contamination source detection in water networks under demand uncertainity

Accidental or intentional drinking water contamination in urban distribution systems may cause great danger for public health. Methods for detecting contamination and for localizing the pollution source are then required. While many recent studies were devoted to the task of quality sensor placement on a water distribution network, the source identification problem was less investigated. Source identification is the inverse problem of water quality simulation in a hydraulic network, which can be solved in direct or indirect ways. An important question is that as many inverse problems it may be ill-posed and then it may present a strong dependence from the input data. This is a crucial point in hydraulic networks, where input data as water demands can be estimated only with an high level of uncertainty. For this reason, only source identification procedures, robust respect to such uncertainty, could be applied to real networks. In this work a procedure able to identify source and input magnitude of an accidental intrusion of pollutant in a hydraulic network using time-varying concentration measurements is presented. The methodology solves the inverse problem in an indirect way, based on a pathway analysis of the network and on the demand coverage concept. In a first step, a subset of candidate nodes than may be potential source of pollutant is individuated. Then, among candidate nodes the source is identified solving a linearized optimization problem, which incorporates a hydraulic network simulation model. Methodology effectiveness is demonstrated through an application to a midsize sample network. Then, a Montecarlo analysis is performed in order to demonstrate the robustness of the proposed methodology respect to nodal demand uncertainty. The analysis results show that the source identification procedure has a reliability around 80% also at high water demand uncertainty levels. Moreover, in the performed tests the error on input concentration estimate is held around 10%. Considerations about the effect of increasing the frequency of measurements in time on the results are presented too.