Deeply flexible commercial building HVAC system control: A physics-aware deep learning-embedded MPC approach

Heating, ventilation, and air conditioning (HVAC) systems within commercial buildings can serve as flexible resources to promote the integration of renewable energy into power systems. However, the complicated operational characteristic of chiller and multi-zone thermal dynamics in the coupled water and air loops lead to a high model complexity to HVAC system control, limiting its operational flexibility exploitation. To tackle this problem, this paper proposes a physics-aware deep learning-embedded model predictive control (MPC) approach to enable deeply flexible commercial building HVAC system control for demand response. Firstly, the chiller's operational characteristic is captured via a deep learning model with high approximation capability, integrated with a physics-constrained block to enforce operational constraints. The multi-zone thermal dynamics are modeled using a graph convolutional network informed by the prior building structure. Secondly, the proposed deep learning models are equivalently reformulated into mixed integer linear constraints and seamlessly embedded into the MPC framework. To enhance the solution efficiency, the bound forward propagation algorithm and network pruning techniques are both developed for the deep learning-embedded MPC approach. Finally, a high-fidelity commercial building HVAC system consisting of coupled water and air loops, as well as outdoor weather conditions, indoor occupancy behaviors, etc. is built on the EnergyPlus simulation program. Comprehensive experimental results have validated the effectiveness of the proposed method in improving flexibility utilization.