TO-22-C029 - Discovering Potential Driving Factors of Smart Thermostat Hold Behaviors Using a Mixture of Logistic Regression Models and Bayesian Inference

Smart thermostats are similar to programmable thermostats in that they allow users to preset temperatures for different times and temporarily override the setpoint temperatures (referred as hold behavior), but they are also equipped with smart features that enable automatic adjustment of setpoint temperature based on occupancy. In addition, some smart thermostats also connect to the Internet and collect and upload detailed indoor environmental data and thermostat setting data. Using such data from the Donate Your Data (DYD) program by ecobee, this paper explored the potential factors that drive users to create a hold. In order to discover the clusters of users having similar behavioral patterns, as well as investigate whether people behave differently in summer and winter, a mixture of logistic regression models was used to find the correlations between the hold behavior and environmental and contextual factors. The factors that highly correlate with the hold behavior were considered to be the potential driving factors. Finally, three clusters were found, and there are two seasonal sub-models within each cluster. The indoor temperature is the most common variable that correlates with both the raising and lowering behaviors, as expected. The signs of the coefficients also align with domain knowledge. In addition, the users are sensitive to over-conditioning in both winter and summer, which suggests that properly guiding users to set their scheduled setpoint temperatures not only improves energy efficiency, but increase the users’ overall satisfaction. But the outdoor temperature and solar radiation contradict the domain knowledge. Further analysis implies that the correlations between the hold behaviors and these two variables could be rooted by a confounder, the time of the day. Therefore, logistic regression is only capable of discovering potential driving factors. The result suggests that future studies need to holistically investigate relationships between variables focusing on causality.