Graph attention neural network for water network partitioning

Abstract Partitioning a water distribution network into several district metered areas is beneficial for its management. according to node features and connections remains challenge. A recent study has realized partitioning based on or pipe individually. This proposes an unsupervised clustering method nodes graph neural network, which uses attention technology update the cluster nodes. The similarity between located in each area balance of total demand are optimized, importance boundary pipes calculated determine installation position flowmeters valves. Three networks with different structures sizes used verify proposed model. results show that average location differences (LocDiffs) within three completed by 0.12, 0.07, 0.06, (DemDiffs) 0.13, 0.27, 0.29, respectively. LocDiff DemDiff decreased 6% 55%, respectively, when compared traditional method. Additionally, calculating boundaries provides objective basis closure. When same number closed, comprehensive impact decreases 17.1%. can be practical applications because it ensures highly reliable interpretive