Model fusion algorithms for digital twinning in built environments: Extending behavioral models in a real HVAC system

Abstract

Digital twin-enabled building operations can improve energy efficiency and reduce carbon emissions from heating, ventilation, and air conditioning (HVAC) systems. However, designing a digital twin model without a measured target variable is inevitable owing to the limited sensing environment and sensor malfunctions in massive building systems. To address this challenge, this study proposes a model fusion method that enables in situ digital twinning for HVAC systems. The proposed method provides an algorithm to effectively combine three model fusion techniques: model coupling, prediction model assembly, and benchmark model assembly, thereby achieving a more accurate and extensive digital twin model environment during HVAC operations. Model coupling is a technique that indirectly calibrates a physics-based prediction model using a benchmark model developed through a data-driven approach. Model assembly involves the additional use of auxiliary models to prevent modeling failures, where prediction model assembly targets the prediction model, and benchmark model assembly focuses on the benchmark model. The proposed method was applied to the (1) mass flow rate and (2) water temperature at the return-side chilled water loop in a real HVAC system. After applying the combinations of model fusion techniques according to the proposed method, the mass flow rate was obtained with a mean absolute percentage error (MAPE) of 2.91 %, and the return water temperature was obtained with a root mean squared error (RMSE) of 0.47 °C. These results demonstrate the effectiveness of model fusion techniques and their combinations for enhancing the accuracy of in situ digital twinning and extending in situ behavioral models for operational HVAC systems.