"The operation of buildings is responsible for approximately 28% of all energy-related CO2 emissions. Despite causing such a large proportion of emissions, most buildings are not operated as well as they could be, resulting in an untapped potential to reduce emissions. Recognising this opportunity, there has been extensive work on improving building controllers using advanced control methods, such as model predictive control (MPC) or reinforcement learning (RL). These state-of-the-art methods have been proven to reduce emissions on test systems. Despite these results, few of these controllers are used outside research settings as most methods require prohibitively expensive expert tuning to be applied to any new building.
To make it easier to deploy more advanced building controllers, we are working on generalist building controllers that can control a wide variety of buildings - with the same controller and no further expert tuning. These controllers have the potential to make many buildings operate a little greener. In this talk, I will outline how we plan to develop such generalist controllers and will present preliminary results. In particular, I will demo our open-source building simulation framework, named Bauwerk, and discuss early results of applying multi-task and meta reinforcement learning (RL) methods to this experimental setup."
