Adaptive threshold optimization for efficient and robust high-dimensional sensor calibration in building energy systems

With the rapid advancement of artificial intelligence and big data technologies, high-quality data have become increasingly essential for effective building energy system management. High-dimensional in-situ calibration provides a promising way to correct sensor biases in building energy systems without relying on physical reference sensors. However, practical applications face significant challenges due to random measurement noise and the computational burden of high-dimensional data. To address these issues, this study proposes a sensor in-situ calibration framework integrated with Bayesian optimization, referred to as SIC-BO. In the proposed framework, a Gaussian probability-based confidence evaluation method is first developed to quantify sensor credibility under physical and semi-physical constraints. Sensors with low confidence are then selected as calibration candidates, thereby reducing calibration dimensionality and improving practical identifiability. Bayesian optimization is further employed to adaptively determine the sensor-confidence threshold and objective-function threshold, so that solution stability and computational efficiency can be balanced without empirical trial. The proposed method was validated using a chiller plant system in a commercial building in Hong Kong. Across five fault scenarios with different fault scales and random-error levels, the selected calibration solutions achieved system-level MAPE values below 2%. These results demonstrate the effectiveness of SIC-BO in improving calibration stability and computational efficiency under different fault and noise conditions within chiller plant systems.