A multi-metric benchmarking framework for clustering algorithms in real water distribution networks

ABSTRACT Multi-metric benchmarking of six clustering algorithms across five real water distribution networks using computational, structural, and multi-criteria ranking metrics to evaluate clustering performance. Clustering of water distribution networks (WDNs) is widely used in applications such as district metered area design, sensor placement, and network management. Previous studies have evaluated clustering algorithms using a limited number of networks or performance metrics, hindering their suitability assessment across different network structures. Unlike previous studies, this study presents a comparative evaluation of six clustering algorithms, namely Louvain, Leiden, K-Means, Spectral Clustering, Hierarchical Agglomerative Clustering (AGNES), and Label Propagation. The algorithms were evaluated using five real-world WDNs ranging from 385 to 5,145 nodes and 513 to 5,767 links, representing a wide range of topological characteristics, with meshedness coefficients ranging from 0.06 to 0.17. All algorithms were executed 40 times to account for stochastic variability and then were evaluated using six computational and structural performance metrics. The results show that Leiden and Louvain consistently achieved the highest overall performance. Label Propagation also demonstrated competitive performance. In contrast, K-means, Spectral Clustering, and AGNES generally required longer runtimes and produced lower-quality partitions. A multi-criteria ranking framework was applied to assess algorithm performance under different operational priorities. The findings suggest that network topology influences clustering outcomes. The proposed benchmarking framework provides practical guidance for selecting clustering algorithms for real WDNs.