Leak localization and maintenance in water distribution networks (WDNs) are essential for reducing water losses and operating costs; however, they usually require extensive monitoring and large datasets. This work proposes a methodology that combines topological sectorization of a hydraulic node network and deep learning techniques to improve leak location by selecting representative nodes to reduce the spatial dimensionality of the WDNs. The network is partitioned using a Spectral Clustering algorithm to identify key nodes based on a weighted criterion that considers pressure variability, flow rate, and proximity to the centroid. Subsequently, a Bidirectional Long Short-Term Memory (Bi-LSTM) neural network classifies the cluster and sub-cluster where a leak occurs, using pressure and flow time series simulated in EPANET. This methodology was validated on the L-Town network, achieving an accuracy of 99.94% for cluster classification and 99.82% for sub-clusters, with a validation loss of 0.024%. During validation with 117 unseen leakage scenarios, the model reached an overall effectiveness of 85%. Moreover, Spectral Clustering outperformed K-Means in preserving physical connectivity. These results confirm the efficiency of the proposed methodology and highlight its potential for application in other hydraulic networks.
Pérez-Sandoval et al. (Sat,) studied this question.