Clustering and Forecasting Model for Water Distribution Systems Analysis

The operation and management of water distribution systems rely on water metering devices installed throughout the network. However, the rapid urban development and population growth and the relatively high cost of meters can hinder water providers from uniformly collecting water demand data across the network for billing, operational, or maintenance purposes. As a result, predictive methods are required to understand consumption patterns in areas with missing or malfunctioning metering infrastructure by utilizing information from similar, measured areas to bridge this gap in demand data. This study presents a two-step method to partition a water network into similar demand areas in order to predict demand conditions in unmeasured regions. The clustering step employs network topology and hydraulic characteristics to minimize demand variation across clusters. Next, a data-driven model is trained using demand data from multiple clusters to forecast demand in an unmetered cluster outside of the train set. Our two-step method reports R² values as high as 0.90 when trained on a set of 10 clusters from a benchmark hydraulic network. The developed predictive model for unmetered areas offers water providers an alternative approach to understanding demand dynamics across the entire network.