Data driven occupancy information for energy simulation and energy use assessment in residential buildings

Occupant's schedules and their energy-use behavior are substantial inputs for building energy simulations and energy management in buildings. In practice, most of the research studies consider default occupant schedules from the standards. Subsequently, the temporal variations associated with occupancy is often missed out, leading to uncertainties in simulation results. This study aims to address two research problems in terms of occupancy: 1) upon the availability of the data, how to systematically extract the different occupant schedules, 2) when the occupancy data is not available, what are the other commonly logged parameters (such as plug load, lighting energy consumption, indoor carbon dioxide (CO2) concentration, and indoor relative humidity data) that shall be used to represent the occupancy in buildings. Regarding the first objective, a generic data-driven framework with the combination of shape-based clustering and change-point detection method is proposed to extract the distinct occupancy in residential buildings in terms of occupant activity schedule and presence probability. To demonstrate the outcomes of the framework, it was applied to the dataset collected from eight apartments located in Lyon, France. The results show the existence of different occupant patterns in buildings with respect to day of the week and season of the year. To achieve the second objective, linear and logistic regression models were developed to represent the occupant activity level and occupant presence/absence state, respectively. The linear regression model results show that among the examined variables, the lighting, and plug load consumption data along with the hour of the day show better prediction results in terms of adjusted R-2 and mean absolute percentage error. For the occupant presence/absence state, the logistic regression model developed using CO2 concentration and plug load energy consumption dataset shows better results in misclassification error, confusion matrix, and receiver operating characteristic curve. (C) 2020 Elsevier Ltd. All rights reserved.