A Bayesian Non-parametric Approach to Water Consumption Clustering

Advanced Metering Infrastructures generate high frequency water consumption data that exhibit substantial gaps and intermittent zero flow periods, posing serious challenges for traditional clustering methods. We propose a Bayesian non parametric framework based on a Dirichlet Process Mixture Model to group users according to their temporal consumption profiles. The model embeds a latent Gaussian variable within a Tobit like structure, allowing a clear separation between consumption propensity and consumption intensity while treating missing observations as latent variables. A scale normalization and relabeling procedure is introduced to ensure parameter identifiability and coherent posterior inference. An application to a real world dataset provided by SMAT in Turin, Italy shows that the proposed approach successfully identifies distinct behavioral patterns even under severe data scarcity, offering a robust tool for water demand management.