Pipe-level leak localization in water distribution networks using pressure residuals and sensor-topology graph learning

ABSTRACT Leak localization at node or district level is often too coarse for field inspection. This study develops a pipe-level framework for water distribution networks. It identifies the leaking pipe and estimates a continuous position within that pipe. Pipe splitting generated 16,000 simulated single-leak samples on the Net2 benchmark network. Early pressure residuals from 11 sensors formed the input. Each sample covered the first 4 h after leak onset. The proposed model combines bidirectional long short-term memory with graph attention version 2. It integrates temporal encoding with graph learning. The sensor graph uses shortest-path distance, topological hop count, and reciprocal path distance. The model achieved 98.09% validation accuracy. This was 7.25 percentage points above the bidirectional temporal baseline. Accuracy rose from 96.72% at a leak coefficient of 0.2 to 99.36% at 0.8. The longest-pipe group had a mean absolute error of 40.50 m. This was 3.81 times the shortest-group error. The matched comparison shows that graph learning improved pipe discrimination. The repair-oriented output can narrow field inspection and guide local confirmation. However, the results use clean simulation data from one network. Field and cross-network validation remain future work.