Towards automated inference of occupant behavioral dynamics using plug-load energy data

When building systems and occupants use energy, they create data—much of it unstructured and characterized as long running time series. Energy data captured at the plug level offers an opportunity not only to analyze highly granular building activities, but also to infer information about the behavior of occupants. Previous work examining occupant behavior typically seeks to understand how individual occupant schedules can be better modeled to improve the efficiency of building system operations, and therefore they treat individual actions as entirely self-contained. However, individual behavior—including that which draws power in buildings—is highly influenced by the inherent spatial and social network structures of occupants. Therefore, understanding the underlying spatio-social occupant network in a commercial building is integral to driving more energy efficient operations. Doing so is challenging since network relationships are highly complex and difficult to directly measure using traditional methods (e.g., surveys) and will require a deep understanding of occupant behavioral patterns. In this paper, we propose an automated methodology for inferring occupant behavioral patterns by classifying raw plug load data ascribed to individual occupants into energy use states. Our method utilizes a Gaussian Mixture Model to model occupant energy use variability and probabilistically classify energy use data into one of three generalized states. We present preliminary results of our classification algorithm using empirical data from a fifty-person commercial office building in San Francisco, California.