Scalable Machine Learning–Based Approach for Pressure Sensor Placement and Burst Detection/Localization in Large Water Distribution Networks

A novel methodology for burst detection/localization and pressure sensor placement in water distribution networks (WDNs) is proposed, by combining machine learning and graph theory algorithms. After a preliminary clustering of the WDN graph through the Girvan-Newman method, the most central nodes in topological terms are selected. A neural network classification model is then trained using pressure data at these nodes with a grouped regularization approach, to define the optimal pressure sensor number and placement. Spatial/temporal uncertainties of WDN hydraulic behavior are considered by simulating several burst/water demand scenarios. Application to a real-world, large-sized WDN in Italy shows outstanding detection/localization test accuracy for all burst entities by relying on a reduced number of optimally defined monitoring points. Furthermore, a conspicuous computational burden reduction was obtained by leveraging topological preselection of a subset of potential most suitable sensor locations making the proposed method very scalable.