Título: From Cumulative Readings to Consumption Patterns: a Bayesian nonparametric model for hourly smart water meter data
Authors: Daniele Poggio (Polito)
Co-authors: Elisa Brussolo (SMAT, Turin), Edoardo Burzio (SMAT, Turin), Gianluca Mastrantonio (Polito)
Language: en

Smart water meters record cumulative volumes hourly. These data show frequent zeros, structured missingness and an unknown number of user profiles. We propose a Bayesian nonparametric model that addresses all three. A latent Tobit layer handles zeros and treats unobserved increments as latent variables constrained by cumulative readings. The mean is decomposed into level, amplitude and a daily shape on the sphere. The shape has a projected normal prior with circular hourly dependence, so users are clustered by routine rather than volume. A Dirichlet process mixture with a point mass for flat users infers the clusters. Applied to SMAT dataset in Turin, the model identifies interpretable daily profiles.