Urban water demand forecasting via artificial neural network models: a case study in Southern Brazil

Brazil’s growing water consumption and the problems arising from its shortage in urban centers are concerning issues. This study performs an exploratory analysis on water demand and applies demand forecast models using artificial neural networks to a city in southern Brazil. For that, twenty-four short-term demand forecast models were proposed for each assessed demand category (residential, commercial, industrial, public and total). The chosen artificial neural network was the Multilayer Perceptron (MLP) with structures of one and two hidden layers, trained using the Backpropagation (BP) and Resilient Backpropagation (RP) methods. The results showed that the RP-trained two-hidden-layer neural networks are more accurate for forecasts in the residential, industrial, public and total categories, while BP-trained networks perform better in the commercial category. The models including both past demands and the other independent variables showed the best results. The most accurate model was obtained in the total water demand category with a 1.59% mean absolute percentage error (MAPE), followed by residential (1.91% MAPE), commercial (1.94% MAPE), industrial (3.02% MAPE) and public (4.57% MAPE) categories.