Smart meters data for modeling and forecasting water demand at the user-level

Abstract

Smart meters installed at the user-level provide a new data source for managing water infrastructure. This research explores the use of machine learning methods, including Random Forests (RFs), Artificial Neural Networks (ANNs), and Support Vector Regression (SVR) to forecast hourly water demand at 90 accounts using smart-metered data. Demands are predicted using lagged demand, seasonality, weather, and household characteristics. Time-series clustering is applied to delineate data based on the time of day and day of the week, which improves model performance. Two modeling approaches are compared. Individual models are developed separately for each meter, and a Group model is trained using a data set of multiple meters. Individual models predict demands at meters in the original data set with lower error than Group models, while the Group model predicts demands at new meters with lower error than Individual models. Results demonstrate that RF and ANN perform better than SVR across all scenarios.