Continuous model calibration framework for smart-building digital twin: A generative model-based approach

Smart building digital twins represent a significant paradigm shift to optimize building operations, thereby reducing their substantial energy consumption and emissions through digitalization. The objective is to virtually replicate existing buildings' static and dynamic aspects, leveraging data, information, and models spanning the entire life cycle. The virtual replica can then be employed for intelligent functions, including real-time monitoring, autonomous control, and proactive decision-making to optimize building operations. To enable proactive decisions, models within the digital twin must continually evolve with changes in the physical building, aligning their outputs with real-time measurements through calibration. This continuous updating requires real-time physical measurements of model inputs. However, challenges arise in the uncertain conditions of buildings marked by sensor absence, malfunctions, and inherent limitations in measuring certain variables. This study introduces a novel calibration framework for physics-based models, addressing the challenges of continuous model calibration in smart-building digital twins while considering the uncertain environment of physical buildings. Within this framework, a novel generative model-based architecture is proposed. This architecture enables a fast and scalable solution while quantifying uncertainty for reliable calibration. Furthermore, a continuous model calibration procedure is presented based on a pre-trained generative calibrator model. A comprehensive evaluation was conducted via a case study employing a building energy model and multiple experiments. The experimental results demonstrated that the proposed framework effectively addresses the challenges of continuous model calibration in smart-building digital twins. The calibrator model accurately quantified uncertainties in its predictions and solved a single calibration problem in an average time of 0.043 second. For facility-level electricity consumption, Coefficient of Variation Root Mean Squared Error (CVRMSE) values of 6.33%, 10.18%, and 10.97% were achieved under conditions of observations without noise or missing data, with noise, and with noise and missing data, respectively. Similarly, for facility-level gas consumption, the corresponding values were 18.75%, 20.53%, and 20.7%. The CVRMSE scores in both cases met the standard hourly thresholds for building energy model calibration.