Artificial Intelligence Prediction of Pipe Breaks in Water Distribution Systems Assisted by Spatial Clustering Analysis and Graph Neural Network

Water distribution systems (WDS) play an essential role in satisfying residential, commercial, and industrial needs, but undesired pipe breaks could result in severe environmental impacts and economic losses. This study investigated machine learning (ML)-based predictions of pipe breaks by innovatively incorporating spatial features for the effects of neighboring areas. Two new methods, including spatial clustering analysis using ArcGIS and graph neural network (GNN) analysis, were integrated with the ML prediction. Among five ML models serving as benchmarks, the highest R2 value of 0.809 was obtained with the random forest (RF) model. Two spatial clustering methods, Hot spot analysis, and Cluster and Outlier analysis from ArcGIS, were deployed to obtain spatial features as the ML model inputs, resulting in an improved model performance with a higher R2 value of 0.824. The GNN-based Unsupervised GraphSAGE was used to gather spatial features and improve the ML prediction to an R2 value of 0.823. While ML is a black-box model, the importance of the spatial features was elucidated through sensitivity analyses. The results of this study have demonstrated the feasibility of utilizing spatial features as the inputs in pipe failure prediction by integrating ML with spatial clustering analysis or deep learning methods.