Short-Term Water Demand Forecasting: A Case Study

Abstract

This paper describes the use of Artificial Neural Networks (ANN) for short-term water demand forecasting. Statistical techniques are used to analyze and identify relevant input variables for the ANN models. A case study is presented in which ANN is used to forecast water demand with a two-day lead time for the Louisville Water Company supply system, Louisville, Kentucky. Previous studies separate the daily water use into non-seasonal use and seasonal use. This paper differs from this method by separating the whole year into summer and winter season. The partition is carried out according to the significantly different climate effects on short-term water demand. In the winter season, climate factors have little effects on water demand. Therefore, the ANN forecast model is simplified by using only historical demand as inputs. The forecasting accuracy reaches 97.21%, which is adequate for water supply system management in the winter. In the summer season, readily available selected climate factors, such as temperature, relative humidity, dew point, wind speed and rainfall are incorporated in ANN models. The difference of weekday and weekend demand is also considered. The optimal summer forecast ANN model uses maximum temperature, relative humidity, rainfall, and historical demand as inputs to forecast the demand in the next two days. The forecasting accuracy is 95.89%. A persistence model is employed in both winter and summer season demand forecast for the purpose of comparison. The comparison of results shows that ANN models demonstrate a strong capability in extracting the nonlinear relations between water demand and climate parameters. The ANN approach simplifies the short-term demand forecasting modeling process significantly compared with conventional regression models, and the forecasting results are promising.