A novel multi-scale hybrid physical and data-driven HVAC system model for optimal demand response control
Model predictive control (MPC) of building HVAC systems plays a vital role in decarbonizing urban energy infrastructure, yet faces two fundamental implementation barriers: the degradation of data-driven model accuracy under small-sample regimes, and the computational intractability of physics-based models in realtime applications. Therefore, we present a multi-scale physics-data fusion (Ms-PDF) modeling framework and demonstrate its effectiveness in variable speed heat pump (VSHP) focused MPC tasks. In Ms-PDF, data-driven models are adopted at the component level and graph theoretic representations are used at the system level to support rapid collection and structuring of operational data. Building on this, a digital twin of the VSHP system is developed, which is integrated with a building thermal model to obtain a control-to-output dynamic thermal response model. Finally, a demand-responsive MPC controller is designed based on the derived thermal response model. Results demonstrate that the proposed Ms-PDF modeling platform achieves an average relative error of only 1.1%, and the digital-twin platform achieves up to 99.3% accuracy. With the developed MPC strategy, energy consumption and comfort-violation rate are reduced by 5.6%~20.7% and 28%~51%, respectively. These findings establish a new paradigm for interpretable hybrid modeling and provide a scalable pathway for deploying MPC in intelligent building systems.