The effect of temporal resolution on the accuracy of forecasting models for total system demand

Abstract

This paper examines the practical implications of using telemetry data at temporal resolutions ranging from finest to coarsest to accurately forecast daily water consumption. The algorithm implemented performs a new fit for every model following a sliding-window approach where new parameters are estimated and forecasts are generated every 24 h. Models with weekly periodic structures are found to more efficiently remove the autocorrelations with respect to models of the daily periodic type. In addition, it is observed that smaller estimation windows positively affect the ability of the models to adapt to sudden changes in the water demand time series. The daily stochastic model structure in combination with the selected estimation characteristics are shown to significantly improve the production estimates for the water utility used as a case study.