Peak daily water demand forecast modeling using artificial neural networks

Peak daily water demand forecasts are required for the cost-effective and sustainable management and expansion of urban water supply infrastructure. This paper compares multiple linear regression, time series analysis, and artificial neural networks (ANNs) as techniques for peak daily summer water demand forecast modeling. Analysis was performed on 10 years of peak daily water demand data and meteorological variables (maximum daily temperature and daily rainfall) for the summer months of May to August of each year for an area of high outdoor water usage in the city of Ottawa, Canada. Thirty-nine multiple linear regression models, nine time series models, and 39 ANN models were developed and their relative performance was compared. The artificial neural network approach is shown to provide a better prediction of peak daily summer water demand than multiple linear regression and time series analysis. The best results were obtained when peak water demand from the previous day, maximum temperature from the current and previous day, and the occurrence/nonoccurrence of rainfall from five days before, were used as input data. It was also found that the peak daily summer water demand is better correlated with the rainfall occurrence rather than the amount of rainfall itself, and that assigning a weighting system to the antecedent days of no rainfall does not result in more accurate models.