Adaptive Forecasting of Hourly Municipal Water Consumption

An adaptive smoothing-filtering approach for on-line forecasting of hourly municipal water use time series is presented. This method is suitable for forecasting an hourly water-consumption time series that is influenced by changing weather conditions and measurement outliers. The proposed seasonal time-series model and adaptive forecasting algorithm can capture both weekday and weekend cycles and produce very accurate forecasts from 1 h to 24 h ahead. The methodology is based on Winters' exponential smoothing, recursive least squares (RLS), and the Kalman filter. The Winters algorithm is useful for recursive updating and extracting time-varying seasonal factors. The deseasonalized residuals are passed on to the RLS and the filter to correct model errors and to whiten the innovations. The on-line adaptive forecasting system also utilizes a data preprocessing procedure to handle measurement outliers, which are caused by data-recording errors and unmodeled disturbances. The validation tests conducted in the present study show that the forecasting system can maintain surprisingly small prediction errors, despite various unmodeled time-varying climatic variabilities.