Abstract
Water Distribution Networks (WDN) are large-scale systems whose management is a complex task, with increasing
environmental and socio-economic implications worldwide, and with a renewed considerable attention by all stakeholders: local
authorities, regulators, environmental groups and the scientific community.
WDM managers need to reliably estimate the water demand in the short-term (typically 1 day ahead), in order to operate their
reservoirs and treatment plants appropriately to meet demand while reducing costs, in particular energy-related costs for caption,
treatment and pumping.
In this paper the authors propose a fully adaptive, data-driven and self-learning approach to forecast short term urban water
demand in two stages: i) identifying and characterizing typical daily consumption patterns (based, at least, on hourly demand
data) and ii) dynamically generating a set of forecasting models for each typical pattern identified at the previous stage. This
schema permits to deal with nonlinear variability of the water demand at different levels, automatically characterizing periodicity
(e.g., seasonality) and behaviour-related differences among different types of days and hour of the day.
The approach has been validated on the urban water demand data acquired through the SCADA system of the Metropolitana
Milanese (MM) partner, the urban water distribution utility in Milan, Italy.
Moreover, the approach has been developed in order to work also at individual (customer) level, exploiting the new available
technological solutions for smart metering (i.e., Automatic Metering Readers, AMRs) now being installed in MM.