Combining clustering and regularised neural network for burst detection and localization and flow/pressure sensor placement in water distribution networks

This paper proposes a novel methodology for simultaneously addressing design of burst detection/localization machine learning algorithm and flow/pressure sensor placement in water distribution networks (WDNs). A preliminary spectral clustering is performed for defining a suitable WDN partitioning, in order to carry out burst localization at cluster level and reduce the computational burden. A grouped regularization approach is used to define the most suitable flow and pressure sensor placement, inter and intra clusters, respectively, and train a neural network classification model, while explicitly considering sensor cost and data redundancy. Several burst (one at a time, at any WDN node) and water demand scenarios are simulated, for a time window of 30 days, for accounting spatial/temporal uncertainties of WDN hydraulic behaviour. The methodology is applied to a real-world WDN in Italy, pre-clustered in four detection areas, showing outstanding detection/localization test accuracy for small (80 % / on average 86.55 %), and especially moderate (100 % / on average 98.43 %) to large (100 % / on average 96.18 %) burst entities, even when relying on a very reduced number of sensors (just five flow meters located on the boundary pipes between detection areas).