Leak Detection Using Regression Trees

Abstract

This work deals with the problem of water demand modeling in big cities. Well-defined model of water demand allows, among others, to detect leaks. Such a model, to be applicable to the problem of leak detection, should take into account weakly, seasonally and other recurrent changes of a water demand structure. Building the model could be specially difficult for a large-scale water supply systems. In this work, to build the water demand model, a method from the artificial intelligence domain was chosen, i.e., application of regression tree was proposed. Regression trees allow modeling, among others, the above-mentioned changing structure of the water demand. The proposed methodology was applied to the real example that concerns a large water distribution network. The obtained results show, that for normal states of the network no false alarms were detected, while in case of leaks they were detected unambiguously. The method has also some other advantages as: easy interpretability of the model, possibility of its modification and tuning.

The research has been partially financed by the European Social Fund, project “Integrated intelligent system for monitoring and management of a water distribution network in the territory of operation of PWiK Ltd. in Rybnik” performed in the framework of the National Programme of Innovation Economy, Action 1.4.