Pipe Failure Prediction in the Water Distribution System Using a Deep Graph Convolutional Network and Temporal Failure Series

Abstract

Ensuring the safety and reliability of the water distribution system (WDS) manifests significant importance for residential, commercial, and industrial needs and may benefit from the structure deterioration models for early warning of water pipe breaks. However, challenges exist in model calibration with limited monitoring data, unseen underground conditions, or high computing requirements. Herein, a novel deep learning-based DeeperGCN framework was proposed to predict pipe failure by cooperating with graph convolutional network (GCN) models for graph processing. The DeeperGCN model achieved much deeper architectures and was designed to utilize spatial and temporal data simultaneously. Two graph representation methods and three GCN models were compared, showing the best predictions with the “Pipe_as_Edge” method and the DeeperGEN model. To identify the priority of pipe maintenance directly, the prediction targets were assigned as a binary classification question to determine break or not over 1-, 3-, and 5-year periods, with prediction accuracies of 96.91, 96.73, and 97.23%, respectively. The issue of data imbalance was observed and addressed through varied evaluation metrics, resulting in the weighted F1 scores >0.96. The DeeperGCN framework demonstrated potential applications in visualizing pipe failure prediction with high accuracies of 97.09, 96.31, and 97.81% across three periods in 2015, for example.