An Incremental Sensor Placement Optimization in a Large Real-World Water System

Abstract

Supplying modern water systems in smart cities requires the ability to monitor water quality in the production plants and the distribution network. Protection against accidental or intentional events is usually based on an Early Warning Detection System (EWDS) to minimize the impact of any contamination [1]. A major issue is positioning online sensors along the water distribution network to ensure the best protection at minimum cost. Several contributions for tackling such problem have been proposed during the last decade, including a scientific challenge comparing 14 methodologies [2] and two reviews of about 150 articles [3,4]. Currently there is no consensus about algorithms or design objectives to use, and such a problem is NP-hard [5,6]. In this paper, a greedy approach is proposed to near-optimally solve the sensor placement problem dealing with large-scale water systems. This methodology is designed to integrate both the specificity of the studied network (expert/prior knowledge) and flexibility related to the uncertainty of the nature, time, and duration of contamination injections. The method is illustrated to minimize the expected fraction of the exposed population on the largest network in France (about 100,000 nodes, 600,000 connections, above 8,000km of pipes). Costly approaches in terms of computation time, such as MIP (Mixed-Integer Programming) and exhaustive search cannot scale to large networks. A sensitivity analysis is presented using the greedy approach on a sub-network which leads to choose the number of sufficient contamination simulations and the concentration threshold used to detect contaminations. The extensive experiments allow us to highlight the effectiveness and the rapidity of the proposed approach.