Predictive Machine Learning approaches for short-term demand forecasting in water distribution networks

In recent years, the increasing availability of high-frequency data and rapid advances in Machine Learning have led to the development of increasingly accurate and adaptive forecasting models capable of dealing with complex time series and integrating additional explanatory variables. In particular, the latest Deep Learning architectures have introduced a paradigm shift in forecasting, allowing simultaneous modeling of nonlinear temporal patterns and the effect of external factors, such as operational variables, weather conditions, and calendar events.Within this fast-evolving scientific and technological context, this thesis aims to analyze and compare different forecasting methods applied to the problem of short-term water demand forecasting, i.e., the short-term prediction of water demand within a real distribution system. First, an overview of the general problem of demand forecasting for time series data is presented: the main aspects of the problem, the critical issues, and the evolution of the methods used in the literature in terms of applied forecasting models are analyzed, starting from traditional statistical approaches and moving on to Machine Learning methods and new-generation Deep Learning architectures. For each class of models, the main characteristics, advantages and limitations are discussed, with emphasis on the management of multivariate time series and associated covariates. Some of these approaches have been applied to a real case study related to a water distribution network, with the aim of obtaining reliable and dynamic forecasts of the inflows to the main stations in the system. This study shows significant differences between the various approaches, highlighting that in most cases more complex models and a Machine/Deep Learning framework lead to tangible improvements, compared to more traditional statistical approaches, in the management of forecasts, even for critical days and holidays. The analysis also highlights the main issues related to data quality, the presence of non-linear seasonal phenomena, and operational variations in plants. Overall, the work demonstrates how the integration of advanced Machine Learning and Deep Learning methodologies can make a significant contribution to improving water network management strategies, providing flexible and potentially useful forecasting tools to support operational decisions aimed at reducing management costs and maintaining system balance, with possible further developments.