An optimal neural network model for daily water demand forecasting

An optimal neural network model for daily water demand forecasting is presented. It has been reported that variation of water demand is related to the weather. A number of researches have shown that the relationships between daily water demand and exogenous variables usually are nonlinear. However, the majority of the short-term water demand forecasting models published have treated the daily water demands as a stochastic time series, and described the relationships by using linear expressions. This study tackles the complexity of the relationship between daily water demands and exogenous variables. In an effort to more effectively forecast the daily water demands, a neuro-genetic algorithm is adopted in this study, which is a combination of the Neural network and the Genetic algorithm. Temperatures, previous day's water demand, sunshine-duration period, and day type have significant impact on the daily water demand forecasting. If only one input parameter is to be used, a model which uses previous day's water demand as an input parameter shows the best results. Among all the models tested in this study, a neuro-genetic model with input parameters of two previous days' demands and today's and yesterday's average temperatures shows the best performance in today's water demand forecasting. It is recommended that a number of models with various input parameters be tested before any particular model is adopted for a specific service area.