TY - JOUR
T1 - Graph attention neural network for water network partitioning
AU - Rong, Kezhen
AU - Fu, Minglei
AU - Huang, Yangyang
AU - Zhang, Ming
AU - Zheng, Lejin
AU - Zheng, Jianfeng
AU - Scholz, Miklas
AU - Yaseen, Zaher Mundher
N1 - Publisher Copyright:
© 2022, The Author(s).
PY - 2023/1
Y1 - 2023/1
N2 - Partitioning a water distribution network into several district metered areas is beneficial for its management. Partitioning a network according to its node features and connections remains a challenge. A recent study has realized water network partitioning based on node features or pipe connections individually. This study proposes an unsupervised clustering method for nodes based on a graph neural network, which uses graph attention technology to update node features based on the connections and a neural network to cluster nodes. The similarity between nodes located in each area and the balance of the total water demand between areas are optimized, and the importance of the boundary pipes is calculated to determine the installation position of flowmeters and valves. Three water distribution networks with different structures and sizes are used to verify the proposed model. The results show that the average location differences (LocDiffs) within the areas of the three networks completed by partitioning are 0.12, 0.07, and 0.06, and the total demand differences (DemDiffs) between areas are 0.13, 0.27, and 0.29, respectively. The LocDiff and DemDiff of the proposed method decreased by 6% and 55%, respectively, when compared to the traditional clustering method. Additionally, the proposed method for calculating the importance of boundaries provides an objective basis for boundary closure. When the same number of boundaries are closed, the comprehensive impact of the proposed method on the pipe network decreases by 17.1%. The proposed method can be used in practical applications because it ensures a highly reliable and interpretive water distribution network partitioning method.
AB - Partitioning a water distribution network into several district metered areas is beneficial for its management. Partitioning a network according to its node features and connections remains a challenge. A recent study has realized water network partitioning based on node features or pipe connections individually. This study proposes an unsupervised clustering method for nodes based on a graph neural network, which uses graph attention technology to update node features based on the connections and a neural network to cluster nodes. The similarity between nodes located in each area and the balance of the total water demand between areas are optimized, and the importance of the boundary pipes is calculated to determine the installation position of flowmeters and valves. Three water distribution networks with different structures and sizes are used to verify the proposed model. The results show that the average location differences (LocDiffs) within the areas of the three networks completed by partitioning are 0.12, 0.07, and 0.06, and the total demand differences (DemDiffs) between areas are 0.13, 0.27, and 0.29, respectively. The LocDiff and DemDiff of the proposed method decreased by 6% and 55%, respectively, when compared to the traditional clustering method. Additionally, the proposed method for calculating the importance of boundaries provides an objective basis for boundary closure. When the same number of boundaries are closed, the comprehensive impact of the proposed method on the pipe network decreases by 17.1%. The proposed method can be used in practical applications because it ensures a highly reliable and interpretive water distribution network partitioning method.
KW - Deep learning
KW - Graph attention
KW - Unsupervised clustering
KW - Water network partitioning
UR - http://www.scopus.com/inward/record.url?scp=85141866453&partnerID=8YFLogxK
U2 - 10.1007/s13201-022-01791-4
DO - 10.1007/s13201-022-01791-4
M3 - Article
AN - SCOPUS:85141866453
SN - 2190-5487
VL - 13
JO - Applied Water Science
JF - Applied Water Science
IS - 1
M1 - 3
ER -