Optimal Sensors Layout for Contamination Source Identification in Water Distribution Systems

Abstract

This paper presents a methodology for optimally allocating sensors for solving the contamination source identification problem in water distribution systems (i.e., finding the optimal layout of a given number of sensors which maximizes the likelihood of contamination source identification). The model is comprised of two stages: at the first stage the water network is divided into influence zones based on the network configuration and hydraulics. Thereafter, all possible combinations of placing the given number of sensors at the different influence zones are tested for their ability to identify contamination sources for a set of pollution events (i.e., a set of contamination injections from different parts of the network, with different injection mass, duration, and starting times). A genetic algorithm framework is used for the contamination source identification. The GA is coupled with EPANET and applied to disclose pollution event characteristics (i.e., injection time, duration, and concentration) using the sensors measured data. The GA fitness function is of a least square type measuring the Euclidean distance between computed and measured concentrations at the sensor locations. The genetic algorithm decision variables (i.e., each genetic algorithm string) incorporate: (1) the contaminant injection node; (2) the injection mass rate; (3) the injection starting time; and (4) the injection duration. The global minimum for a least square minimization problem is known (i.e., zero) and thus when obtained, the contamination source is identified. At the second stage the combination that maximizes the contamination source identification likelihood is used as the search space at which only nodes from this combination can be selected as possible sensors locations. The result of the two search processes is the optimal sensors layout that maximizes the contamination source identification likelihood. The effectiveness of the method is demonstrated through two example applications.