Adaptive Partition Optimization Method for Water Distribution Network Monitoring Points Driven by Multi-Source Data Fusion

This study proposes a multi-source data fusion method for adaptive monitoring partitioning and sensor placement in water distribution networks (WDNs). It aims to improve leakage-sensing coverage and regional localization accuracy with limited monitoring resources. With boundary-node assignment (BNA) adjustment as its focus, the method first integrates nodal pressure and pipe flow time series under normal operating conditions with pipe diameters, pipe lengths, and network topology. These inputs form a hydraulically weighted graph that couples network topology with hydraulic characteristics. Candidate partitioning schemes are evaluated to adaptively determine the number of monitoring partitions. Graph learning and connectivity constraints then generate spatially continuous initial partitions. BNA adjustment and sensitive sensor selection within each partition are formulated as dual optimization objectives. Partition boundaries and sensor locations are iteratively updated using a sparse-matrix-based coverage criterion to obtain the adaptive monitoring scheme. The method was evaluated using the Wanfudong WDN. Network topology, pipe diameters and lengths, and nodal demands were obtained from field survey data. Hydraulic simulations using the Environmental Protection Agency Network Analysis Tool (EPANET) generated pressure, flow, and leakage-response data for evaluating network partitioning, sensor placement, and leakage region identification. The proposed method was compared with K-means, density-based spatial clustering of applications with noise (DBSCAN), and the Structural Deep Clustering Network (SDCN), with 13 pressure-monitoring points per scheme. Before joint optimization, the initial monitoring scheme constructed using nodal hydraulic features and network topology achieved 89.42% leakage coverage at a leak intensity of 0.8 under the design conditions. This coverage exceeded that of all three baselines: K-means, DBSCAN, and SDCN. Joint refinement of boundary-node assignments and sensor locations further increased coverage to 94.43%, a gain of 5.01 percentage points. Using the optimized sensor configuration, a multiclass support vector machine (M-SVM) achieved a leakage region identification accuracy of 97.46%. These results indicate improved leakage-sensing coverage and regional localization accuracy for the simulated leakage scenarios.