Nonintrusive in-situ modeling for unobserved virtual models in digital twin-enabled building HVAC systems: A one-year comparison of data-driven and physics-based approaches in a living laboratory

Virtual modeling of building systems presents significant challenges due to the prolonged operation under dynamic conditions. This study proposes a novel, nonintrusive in-situ modeling method to overcome the practical measurement difficulties while providing reliable long-term operational data. Previous research has achieved high-accuracy virtual models; however, their applicability was often limited to specific time periods or singular operating conditions. Building on these efforts, this study establishes an integrated framework that incorporates multiple virtual models developed through a sequence of prediction, benchmarking, and correction processes. To optimize model performance, data-driven and physics-based modeling approaches were employed for comparison. Field studies were conducted on a real heating, ventilation, and air conditioning system (HVAC), and long-term modeling performance was validated over a one-year period. The physics-based corrected model demonstrated superior performance, with the annual root mean squared error (RMSE) value significantly reduced from 0.592 m³/h to 0.034 m³/h, and the mean absolute percentage error (MAPE) decreased by 26.7 %. In summary, physics-based nonintrusive in-situ models offer excellent long-term applicability and facilitate optimal control and holistic monitoring of building system operations.