Grant: Discovery Launch Supplement

The electrical grid is facing challenges due to increased electricity demand and the move towards electrification in various sectors. For instance, buildings, already accounting for 60% of electricity use in Canada, are projected to experience a further 52% surge by the year 2050. However, buildings could also play a pivotal role in addressing these challenges by better managing their electricity use, or more generally, upgrading to grid-interactive buildings. To equip buildings with grid interaction capabilities, advanced controls such as Model Predictive Control (MPC) are an essential component, as the current building control system is not designed for such purpose. Yet, field studies have shown some of the limitations of MPC, such as a lack of adaptations to operation changes. Machine Learning (ML) is a promising technique to complement the conventional MPC on those aspects, for instance, ML could be incorporated into MPC to consider uncertainties due to evolving building occupancy. Despite a few early attempts recently, this hybrid approach of MPC and ML is still nascent in the building sector. In addition, its application on grid-interactive buildings is very rare. This program therefore aims to investigate a systematic approach to combining MPC and ML within a unified framework, termed learning-centric MPCs, with a focus on their application to grid-interactive buildings. The long-term goal of this program is to enable buildings to provide energy flexibility services to the grid. This helps to enhance grid resilience during peak hours and boost energy efficiency at the grid level. It also allows the grid to integrate more renewables and achieve reductions in greenhouse gas emissions. The proposed learning-centric MPCs will act as the brain for grid-interactive buildings. It sends optimal setpoints of temperature or power based on its predictions to controlled energy systems in buildings. As a result, buildings will be operated with a higher energy performance with improved thermal comfort, and in tandem with the grid as well. The proposed control framework will be validated and tested virtually as well as in real buildings. To this end, virtual testbeds based on high-fidelity models will be developed. The advantage of virtual testbeds is that they allow fair comparisons of different ML algorithms in the same environment. On the other hand, real-world implementation helps to validate the applicability of the proposed approach and identify barriers to large-scale adoptions. This program will train 3 Ph.D., 2 M.A.Sc. and 3 B.Eng. The knowledge transfer to the industry will in the long term allow Canada to continue being a global leader in the clean energy sector. The transformation to grid-interactive buildings promises to enhance energy efficiency, increase renewables integration, curtail greenhouse gas emissions and ultimately help Canada achieve its carbon neutrality goal by 2050.