Observation and prediction models play important roles in intelligent building systems in terms of holistic monitoring and optimal control. These models overcome the limitations of physical sensors and contribute to organizing the building sensing environment into a more informative arrangement. However, developing an in-situ prediction model for real systems is challenging, as the target sensor observation (Y) is absent. Although physics-based white-box modeling is possible, the model observations have operational uncertainties in a real system. Therefore, this study proposes a novel and integrated framework for a nonintrusive high-accuracy observation modeling method including (1) white-box modeling, (2) indirect-nonintrusive calibration, and (3) surrogate-based uncertainty reduction. In contrast to the conventional application purpose of a surrogate model, the proposed method adopts surrogate modeling to reduce the model uncertainty and provide high accuracy. The proposed method is applied to a real building system to evaluate the accuracy of each modeling step in the proposed framework. The initial white-box model shows an error of 0.81 °C, indirect-nonintrusive calibration improves the accuracy with an error of 0.55 °C, and the surrogate-assisted model reaches an error of 0.27 °C. Notably, 51% of sensor errors are decreased through surrogate modeling.
Read full abstract