Abstract
Sustainable management is the keyword for the future. The constant increase of water demands due to socio-economic factors and climate change are putting under an increasing strain our distribution systems. In this condition, water management needs proper care to undertake a transition towards a smart and sustainable paradigm. Researcher nowadays are dealing with this need. The new era of Big data and artificial intelligence allows innovative ways to make the water management more efficient. For instance, it is possible to develop reliable data-driven methods to predict water demands, to detect anomalies in the water systems, to find optimal locations for sensors, and much more. However, this is a fast-changing world, that keeps updating with newer and more performing methodologies year after year. In this context, hydroinformatics research plays a crucial role. This thesis aims at giving a contribution towards the future management of the water resource, by developing novel and performing water demand forecasting methodologies. More in detail, this thesis deals with the challenge of developing an innovative water demand forecasting model, and with some related problems. These latter include the study of data generation methods to create realistic water demand time series for testing algorithms, the study of data imputation methods and their influence with some forecasting methodologies, and also the study on tuning process for neural network model aimed at forecasting purposes. At the end of this manuscript, this thesis proposes two novel approaches to forecast water demands based on two different ideas: an ensemble and a graph-based approach. The main outcomes of this thesis support the novel forecasting methodologies and also proposes novel state-of-the-art methods. On the one hand, the studies on data generation, on data imputation and on the tuning provide new considerations on how to develop these data-driven methods, but also dig into some aspects of the crucial role played by data in the nowadays methodologies. On the other hand, the novel forecasting methodologies aim at providing new and innovative ways to accomplish robust and reliable prediction of water demand, also in context with strong variability of the data, or in context with sensor malfunction.