Author(s): Biniam Abrha Tsegay; Ali Khajavian; Nicolas Peleato

Linked Author(s):

Keywords: Data augmentation; Generative modeling; Smart water meters; Water demand prediction

Abstract: Urban Water Distribution Systems (WDS) require accurate demand forecasting for effective planning and resource management. However, traditional forecasting models often fail under different environmental changes that alter consumption patterns beyond their training scope. This study proposes a novel model, Environment Aware Time-Variational Autoencoder (EA-TimeVAE), a generative framework that produces realistic, weather-conditioned synthetic water demand data to enhance forecasting robustness. Validated on the FP7 DAIAD dataset (Alicante, Spain; 1,099 users), the proposed augmentation consistently reduces forecasting error across multiple model architectures and aggregation levels. Performance gains increase systematically with aggregation, exceeding 60% Normalized Root Mean Square Error (NRMSE) reduction at higher spatial scales. Robustness analyses across randomized trials further show that these improvements are driven by data characteristics rather than model-specific behavior, indicating that EATimeVAE provides a model-agnostic and reliable enhancement strategy for urban water demand forecasting.

DOI: https://doi.org/10.64697/iahr.proc.hic2026.65

Year: 2026