A fully adaptive forecasting model for short-term drinking water demand

• Our model forecasts the water demand in 15-min time steps for the next 48 h.
• The model discerns patterns for the days of the week and a number of deviant days.
• We tested the model on six datasets of six years of water demand.
• The average forecasting errors varied between 1.7 and 5.5% for the 24-h forecasts.
• And between 3.5 and 10.6% for the 15-min forecasts.

For the optimal control of a water supply system, a short-term water demand forecast is necessary. We developed a model that forecasts the water demand for the next 48 h with 15-min time steps. The model uses measured water demands and static calendar data as single input. Based on this input, the model fully adaptively derives day factors and daily demand patterns for the seven days of the week, and for a configurable number of deviant day types. Although not using weather data as input, the model is able to identify occasional extra water demand in the evening during fair weather periods, and to adjust the forecast accordingly.The model was tested on datasets containing six years of water demand data in six different areas in the central and Southern part of Netherlands. The areas have all the same moderate weather conditions, and vary in size from very large (950,000 inhabitants) to small (2400 inhabitants). The mean absolute percentage error (MAPE) for the 24-h forecasts varied between 1.44 and 5.12%, and for the 15-min time step forecasts between 3.35 and 10.44%. The model is easy to implement, fully adaptive and accurate, which makes it suitable for application in real time control.