A Statistical Approach for Unsupervised Occupancy Detection and Estimation in Smart Buildings

The energy usage of a building depends significantly on the number of occupants inside. Therefore, occupancy detection and estimation are crucial for efficient energy consumption planning. These two tasks have been generally tackled using supervised machine learning techniques. Unlike these previous efforts, the aforementioned tasks are carried out, in this paper, automatically in unsupervised settings using a statistical framework based on finite mixture models. The main idea is based on modeling sensor features as a weighted sum of probability density functions. Unlike previous approaches in mixture modeling literatures that have generally considered Gaussian distributions, we consider scaled Dirichlet distribution that has shown recently great flexibility and efficiency in various challenging applications. In particular, we propose a novel algorithm to learn finite scaled Dirichlet mixture models via an entropy-based variational Bayesian inference approach. The results of the proposed framework are analyzed taking into account comparable methods in order to validate its efficiency.