CPS: Medium: A meta-learning approach to enable autonomous buildings

Buildings are vitally important because they contribute to the well-being and productivity of their occupants - however, these benefits come at a high environmental cost. Collectively, buildings account for 40% of the US primary energy usage and CO2 emissions and 70% of the electricity consumption. Furthermore, buildings put a tremendous strain on the power grid as they are largely responsible for the peaks in energy demand.

Making buildings smarter through the deployment of sensors, actuators, and controllers, which collectively serve as the backbone of building cyber-physical systems (CPS), can achieve more than 30% annual energy savings and can also significantly smooth peak demand. Thus, smart buildings are vital to a sustainable energy future. However, the road to large-scale realization of smart buildings is inhibited by their heterogeneity, which requires engineering customized, site-specific, and, thereby, costly solutions.

The goal of this project is to develop a CPS solution for autonomous buildings that will enable non-expert building managers to deploy asset-specific, smart control policies. The advantage of the proposed solution relies on the fact that the approach can be applied on a large-scale even without any human intervention. The resulting software solution is the Artificial-Intelligence-Enabled Building Energy Expert (AI-BEE) and it will be demonstrated using simulations and experiments at the Center for High Performance Buildings at Purdue University.

The proposed research will result in foundational contributions in core CPS areas, including machine learning and control, that will be translational to other application areas, such as large-scale energy systems (power grid), transportation, civil infrastructure, and unmanned vehicles.

The technical details of our approach are as follows. First, a taxonomy of building types is being developed. The idea is that the energy behavior of every building should be completely specified by a finite set of variables in a machine-readable format.

Second, each complete building description is associated with a set of dynamical systems that describes the energy consumption. In this way, non-experts will be able to specify building characteristics and get a set of plausible dynamical systems that include a description of the building. This set of dynamical systems is what is called the relevant model universe to the building at hand.

Third, meta reinforcement learning is being used to discover a self-improving control algorithm that works well for all dynamical models in the relevant model universe. The final step is to deploy the discovered algorithm to the building and let it self-improve further.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.