Enhancing the Detection of Contamination Events in Water Distribution Systems by Integrating Alternative Sources of Information in Real-Time

Abstract

Water distribution systems (WDS) are one of the most vulnerable components of water infrastructure systems. After the events of 9/11, there was increased concern regarding the potential for the deliberate introduction of chemical agents that may affect public health. Later, these concerns moved towards accidental water quality incidents produced by, for example, cross-connections and low-pressure events. In either case, Surveillance Response Systems (SRS) were formulated to protect water quality by make prompt detection and confirmation of contamination events. The overall objective of this dissertation is to develop a decision framework capable of making a timely detection, confirmation and forecasting of a contamination event in WDS. The fulfilment of this objective relied on the use of confirmatory sampling locations (CSLs) as a way to overcome the limitations of traditional water quality monitoring, which includes the high cost of sensors and limited detection capabilities due to the lack of contaminant-specific sensors. Based on previous research that provided a foundation to identify the best individual CSL using an entropy metric derived from information theory, this dissertation proposed the Updating Greedy Algorithm (UGA) to identify a set of \emph{multiple} CSLs in real-time. The UGA used an approximation to the entropy metric and a greedy heuristic to identify the set of CSLs. When tested on a small network, the UGA results were almost identical to the solutions achieved by a Genetic Algorithm using an exact formulation of the entropy metric, but in orders of magnitude less time. As the number of CSLs increased, the placements were spatially very similar (similar hydraulic paths), which may lead to a potential overlap of information across the CSL. To enable the implementation of the UGA in real-sized networks, the UGA was extended into a cluster-based optimization approach in order to reduce the solution space. The clustering algorithm grouped the network nodes based upon similarities in hydraulic connectivity. Each cluster was then assigned a representative sampling unit within each cluster to be used as a potential CSL. Two types of sampling units were used: i) the centroids of the cluster and ii) the location within the cluster that maximized information gain. The use of the clusters and sampling units were proposed to reduce the computational burden of the algorithm while maintaining a similar set of spatially distributed potential CSL. Ultimately, this approach was tested on a large network, and was shown capable of generating optimal solutions within an amount of time consistent with real-time application (1 hour). Finally, the reduction of information overlap was investigated through the introduction of the Multi-Objective Updating Greedy Algorithm (MUGA). The MUGA incorporated D-optimality, a correlation metric, to increase the spatial distribution of CSLs across the network while maximizing the amount of information gained from the samples. The trade-off between Information Gain and D-Optimality showed that the information gain tended to place CSLs along similar hydraulic paths, reinforcing existing connections, whereas the maximization of D-Optimality placed CSLs in unexplored hydraulic paths, creating new connections. The MUGA was tested in a large network using the clustering approach previously developed. The MUGA approach was shown to be computational efficient, and able to achieve good results (relative to Genetic Algorithm solutions). Overall, the addition of the D-optimality was capable of spatially distribution the CSLs and, at lower weights, was able to help avoid some local minima when using the Information Gain as a single objective. Overall, the deployment of CSLs provides useful information to characterize a contamination event, that enhances detection, confirmation and forecasting capabilities of SRS. These methodologies must be efficient enough to provide good solutions in at a time frame useful for real-time application. This research developed an approximation to the exact solution for information gain, a clustering approach to reduce the solution space of large networks, a greedy heuristic placement algorithm, and a multi-objective placement algorithm that that were shown capable of generating good CSL solutions in less than an hour -- often in seconds for the single objective placement. Further research may explore the response of the proposed approaches under diverse characteristics of a contamination event (varying injection locations and times), as well as considering demand uncertainty. Finally, as some contaminants may remain undetected, the inclusion of alternate data sources (e.g., public health data) may be integrated to enhance detection capabilities of SRS.