Save Money or Feel Cozy?: A Field Experiment Evaluation of a Smart Thermostat that Learns Heating Preferences

We present the design of a fully autonomous smart thermostat that supports end-users in managing their heating preferences in a real-time pricing regime. The thermostat uses a machine learning algorithm to learn how a user wants to trade off comfort versus cost. We evaluate the thermostat in a field experiment in the UK involving 30 users over a period of 30 days. We make two main contributions. First, we study whether our smart thermostat enables end-users to handle real-time prices, and in particular, whether machine learning can help them. We find that the users trust the system and that they can successfully express their preferences; overall, the smart thermostat enables the users to manage their heating given real-time prices. Moreover, our machine learning-based thermostats outperform a baseline without machine learning in terms of usability. Second, we present a quantitative analysis of the users' economic behavior, including their reaction to price changes, their price sensitivity, and their comfort-cost trade-offs. We find a wide variety regarding the users' willingness to make trade-offs. But in aggregate, the users' settings enabled a large amount of demand response, reducing the average energy consumption during peak hours by 38%.