Buildings account for nearly 40% of the total energy consumption in the United States, about half of which is used by the HVAC (heating, ventilation, and air conditioning) system. Intelligent scheduling of building HVAC system has the potential to significantly reduce the energy cost as well as to improve occupant's comfort. For instance, if we know occupants' behavior and schedules we can turn off the HVAC system when not needed or customize them to meet occupants' preferences. However, the rule-based and other traditional pre-programmed controllers are ineffective in this regard, mainly due to stochasticity in the system (e.g. occupancy of the rooms and weather conditions) and huge differences in buildings' dynamics.
With recent advances in sensor technology and internet of things, smart buildings have received much attention in the recent years. Smart buildings use information and communication technologies to enable automated building operations and control to enhance occupants' comfort and reduce energy consumption. The current smart technology operates and makes decisions mainly based on occupancy predictions. However, real-time weather data as well as coordination between HVAC devices could be as important in the decision process. In this project, we are developing a framework to learn effective policies for decentralized and coordinated climate control of buildings based on some relevant real-time and historical data such as weather conditions, user-defined preferences and settings, sensory signals of the inside climate such as temperature, light intensity, humidity, and CO concentration, signals from the HVAC devices, and occupancy condition and activities. To this end, we are tailoring reinforcement learning techniques by integrating physical laws and constraints into these techniques to create a more structured learning problem; this results in significant reduction in data requirements (currently the biggest limitation of applying these techniques to physical problems) and physically plausible and reliable solutions.