An iterative method for leakage zone identification in water distribution networks based on machine learning

Abstract

For leakage identification in water distribution networks, if each node is used as a category label of the classifier model, the accuracy of the classifier model will be low because of similar leakage characteristics. By clustering the nodes with similar leakage characteristics and using all the possible combinations of leakages as the category labels of the classifier model, the accuracy of the classifier model for leakage location can be improved. An iterative method combining k-means clustering with the random forest classifier is proposed to identify the leakage zones. In each iteration, k-means clustering is used to divide the leakage zone identified in the previous iterations into two zones, and then, the random forest classifier is used to identify the leakage zones and the number of leakages in each leakage zone. As the number of iterations increases, the number of candidate leakage zones and sensors that conduct leakage zone identification decreases. Thus, feature selection can be used in each iteration to select the minimum number of sensors for model training without affecting identification accuracy. Three leakage scenarios are considered: a single leakage, two simultaneous leakages, and four simultaneous leakages. A benchmark case is presented in this study to demonstrate the effectiveness of the proposed method. The influences of the number of pressure sensors and Gaussian noise level on the identification results are also discussed. Results indicate that the proposed method is effective for identifying simultaneous leakages.

An innovative machine learning based framework for water distribution network leakage detection and localization

“…Zhang 24 and Chen 25 defined the leakage characteristics matrix using the change of monitored water pressure due to a given leakage occurring at each junction compared with non-leaking conditions. Table 1 illustrates the calculation of the leakage characteristics matrix, where each row is the leakage characteristics vector corresponding to the junction.…”

“…The standard k-means algorithm has been used in previous studies for the WDN zone partition to reduce the degree of freedoms in leakage detection, based on the conventional leakage characteristics matrix. 24,25 The partition aims to improve leakage detection and localization accuracy.…”

“…Leaking is assumed to occur at each junction, which is convenient for clustering purposes. 24,25 The leakage size is assumed to be randomly set between 0.05 m to 0.1 m in this study, which leads to 0.0012m2 to 0.0078 m2 leaking size. These, however, can be easily changed to more complex leaking conditions.…”

“…Recent studies also explored leakage detection by clustering junctions into multiple leakage zones to enhance the final detection accuracy. 24,25 Although supervised learning approach can achieve a high leakage localization accuracy, it requires a balanced dataset, which means it needs a sufficient amount of WDN operational data under both leaking conditions and non-leaking conditions. However, as pointed by Mounce et al, 26 datasets under leaking conditions are very scarce.…”

A flexible, leak crew focused localization model using a maximum coverage search area algorithm

“…A few leak location models have been designed to predict a zone within the network that is likely to contain the leak. The leak zone prediction models have been developed in two distinct ways; through classification of predefined leak zones / district metered areas [9], [10], or through iterative cluster / graph theory partitioning [11], [12]. By splitting up the network into predefined leak zones, these location models predict which zone the leak is occurring in.…”