FORECASTING HOURLY WATER DEMANDS BY PATTERN-RECOGNITION APPROACH

Hourly water-demand data is forecasted with a model based on a combina tion of pattern recognition and time-series analysis. Three repeating segments are observed in the daily demand pattern: ''rising,'' ''oscil lating,'' ''falling,'' then ''rising'' again the following day. These are called ''states'' of the demand curve, and are defined as successi ve states of a Markov process. The transition probabilities between st ates are ''learned,'' and low-order auto-regressive integrated moving average (ARIMA) models fitted to each segment, using a modest amount o f historical data. The model is then used to forecast hourly demands f or a period of one to several days ahead. The forecast can be performe d in real time, on a personal computer, with low computational require ments, at any time the system state deviates from the planned, or when new data become available. The process of model development, applicat ion, and evaluation is demonstrated on a water system in Israel.