Abstract
A reliable short-term forecasting model is fundamental to manage a water distribution system properly. In the latest years, a plethora of methods have been proposed by the scientific community, ranging from classic naïve methods to the more sophisticated statistical and machine learning models. Among several methods, neural networks gained particular attention, especially with the rise of the deep learning approach. Nonetheless, the forecasting of urban water demand is nowadays a demanding challenge for the scientific community, being still a florid research topic year after year. This study addresses the problem of developing a reliable model for short-term forecasting of water consumption of alpine water supply systems. These aqueducts have typically high pressures due to the mountain environment that causes significant differences in the elevations along the distribution network. For this reason, water consumption is affected not only by the users demands but also by a significant component of background leakages and bursts. This study analyzes the potential of recurrent neural network algorithms to establish a reliable and robust forecasting model for these type of consumption. In particular, it is proposed to assess the influence of using different sets of inputs for the model, analyzing many combinations of past observation, calendar variables and also meteorological variables. The results show that the deep learning model based on the recurrent neural network can provide an effective solution to forecast the consumption of mountain aqueducts. Additionally, it is highlighted the crucial role of including the correct past observation in the input for the model.