Clustering-Learning Approach to the Localization of Leaks in Water Distribution Networks

Abstract

Leak detection and localization in water distribution networks (WDNs) is of great significance for water utilities. This paper proposes a leak localization method that requires hydraulic measurements and structural information of the network. It is composed by an image encoding procedure and a recursive clustering/learning approach. Image encoding is carried out using Gramian angular field (GAF) on pressure measurements to obtain images for the learning phase (for all possible leak scenarios). The recursive clustering/learning approach divides the considered region of the network into two sets of nodes using graph agglomerative clustering (GAC) and trains a deep neural network (DNN) to discern the location of each leak between the two possible clusters, using each one of them as inputs to future iterations of the process. The achieved set of DNNs is hierarchically organized to generate a classification tree. Actual measurements from a leak event occurred in a real network are used to assess the approach, comparing its performance with another state-of-the-art technique, and demonstrating the capability of the method to regulate the area of localization depending on the depth of the route through the tree.