Robust leak localization in water distribution networks using computational intelligence

Abstract

The search for new strategies for leak detection, estimation and localization in Water Distributions Networks (WDNs) is a state-of-the-art research topic. In this paper, a methodology for leak detection, estimation and location that combines data-driven and model-based methods is proposed. A deep neural network is used in the leak detection task. Subsequently, the estimation of a leakage size range is accomplished by using Gaussian process regression. Then, a novel approach based on the solution of an inverse problem is developed for leak location. Knowing the range of possible values for the leak size allows to improve the location task when solved as an inverse problem. The proposed location method considers the topological configuration of the network as well as the leak size range. One of the main advantages of the proposal is that it does not depend on the labeling of the nodes. In this sense, a modified variant of the Differential Evolution algorithm, which considers the topological structure of the WDN to modify the search space and incorporates a temporal analysis, is used to find the solution of the inverse problem. Moreover, thanks to the topological evolution of the solutions a set of candidate nodes for the leakage creates a zone of reduced possible locations very useful in practical terms. The proposed approach is tested with the model of a real case study: the large-scale Modena WDN. The results demonstrate the effectiveness of the proposal with satisfactory leak detection, leak size estimation, and location performance when considering only 9 sensors installed in a network formed by 268 nodes.