Development of a Data-Driven Estimator from Smart Meter Data for Total Water Demand Reconstruction

Smart water meters provide detailed measurements of water consumption and enable new opportunities for data-driven modeling in water distribution networks. Unlike traditional meters, smart meters allow utilities to monitor both users and the network at a fine temporal resolution, enabling the early detection of anomalies and abnormal consumption patterns. However, in many districts only a subset of users is monitored by smart meters, while the total inflow is measured at the district level. Estimating the demand of unmonitored users from the available smart meter data therefore represents an important challenge. This work develops a data-driven estimator for hourly water consumption based on smart meter observations. A preprocessing pipeline is applied to real smart meter data. The strong skewness of water demand is taken into account through a logarithmic transformation of the target variable, and the systematic bias introduced by the back-transformation is explicitly addressed. In addition, a specialized model architecture is designed to improve prediction performance for underrepresented user categories. The results show that the trained models provide reliable aggregated hourly estimates, in particular for residential users. As a preliminary analysis for district demand reconstruction, the district inflow measurements are first preprocessed and cleaned. The predicted demand curve largely captures the main temporal dynamics of the observed one, although there is a substantial gap between predicted and measured volumes, indicating the need for further investigation.