A Learning-Based Model Predictive Control Method for Unlocking the Potential of Building Energy Flexibility

Abstract

Buildings are regarded as promising energy flexibility resources due to their significant energy consumption and better integration into the electricity grid. To fully exploit the potential of building flexibility, optimized operational strategies need to be developed in which cost savings and thermal comfort should be considered. Model predictive control (MPC) is widely acknowledged as one of the effective methods for developing optimal strategies. However, the practical implementation of MPC is challenging due to the significant computational resources required for solving online optimization problems. Therefore, this study proposes a learning-based model predictive control (LBMPC) method that employs machine learning models to imitate the behavior of MPC, enabling the direct output of optimal strategies in real-time control, thereby addressing computational burden limitations. The proposed method consists of three parts: identifying a control-oriented building thermal model to formulate the MPC problem, conducting offline simulations to generate the training dataset, and training classification and regression trees (CART) to learn optimal control actions from dataset. The control performance of the proposed method in a multi-zone office building was evaluated using a high-fidelity co-simulation testbed constructed with Modelica and Spawn of EnergyPlus. The results indicate that, compared to the baseline control strategy, traditional MPC and LBMPC reduced energy costs by 13.89% and 12.27%, respectively, and peak electrical loads by 30.35% and 24.78%, respectively, without compromising thermal comfort. Especially, compared to traditional MPC, LBMPC can significantly reduce the computational cost by as much as 99.6%, with only a small trade-off in performance. Besides, the impacts of input features, meteorological conditions, and model accuracy on the control performance of the proposed method are discussed in detail.

Keywords: Energy flexibility, Model predictive control, Machine Learning, Building energy and control simulation, HVAC