Title:Modularized Neural Network Incorporating Physical Priors for Smart Building Control, Accuracy or Consistency?

Abstract:Model predictive control can achieve significant energy savings, offer grid flexibility, and mitigate carbon emissions. However, the challenge of identifying individual control-oriented building dynamic models limits large-scale real-world applications. To address this issue, this study proposed a Modularized Neural Network Incorporating Physical Priors (ModNN), capable of establishing a control-oriented and physical-consistent building dynamic model within minutes without substantial modeling effort. This is also the first study to evaluate the physical consistency of a given data-driven model both qualitatively and quantitively. We compared the physical consistency of a classical Long Short-Term Memory (LSTM) model and our ModNN. The ModNN strictly satisfies physical constraints, whereas the LSTM model learned contradictory system dynamics. Additionally, we compared their control performance on an EnergyPlus virtual testbed. While the LSTM model demonstrated slightly better prediction accuracy in dynamic modeling, it failed in control optimization, resulting in an 89°C-h temperature violation, whereas the ModNN showed only a 0.57°C-h violation and achieved up to a 78% peak load reduction. Our findings highlight the importance of incorporating physics priors into data-driven models and provide a promising solution for future smart building control optimization. Furthermore, the proposed evaluation framework defines two physical consistency indicators, providing guidelines for selecting and testing control-oriented, data-driven building dynamic models.