Sonta, Andrew J.Simmons, Perry E.Jain, Rishee K.2022-11-142022-11-142022-11-14201810.1016/j.aei.2018.04.009https://infoscience.epfl.ch/handle/20.500.14299/192263Buildings are our homes and our workplaces. They directly affect our well-being, and they impact the natural global environment primarily through the energy they consume. Understanding the behavior of occupants in buildings has vital implications for improving the energy efficiency of building systems and for providing knowledge to designers about how occupants will utilize the spaces they create. However, current methods for inferring building occupant activity patterns are limited in two primary areas: First, they lack adaptability to new spaces and scalability to larger spaces due to the time and cost intensity of collecting ground truth data for training the embedded algorithms. Second, they do not incorporate explicit knowledge about occupant dynamics in their implementation, limiting their ability to uncover deep insights about activity patterns in the data. In this paper, we develop a methodology for classifying occupant activity patterns from plug load sensor data at the desk level. Our method makes us of a common unsupervised learning algorithm—the Gaussian mixture model—and, in addition, it incorporates explicit knowledge about occupant presence and absence in order to preserve adaptability and effectiveness. We validate our method using a pilot study in an academic office building and demonstrate its potential for scalability through a case study of an open-office building in San Francisco, CA. Our method offers key insights into spatially and temporally granular occupancy states and space utilization that could not otherwise be obtained.text::journal::journal article::research article