In-situ sensor virtualization and calibration in building systems

In this study, in-situ sensor virtualization is proposed to overcome physical sensing limitations, including sensor errors and management costs in long-term building operations. Virtualization is achieved by a backup virtual sensor for a target sensor when the physical sensor is faulty during operation. This requires in-situ virtual sensor modeling and a continuous calibration process for effective virtual sensing in operational building energy systems. The main research question is how the in-situ calibration for a virtualized sensor can be formulated without a physical counterpart/benchmark. The proposed virtualization with in-situ calibration was applied to central heating systems serving university campus buildings with real operational datasets obtained by a building automation system. The case study showed (1) the effectiveness of virtualization when the target physical sensor has absolute bias errors more than approximately 1.0 °C, (2) the virtualized sensor errors (the average root mean square error (RMSE) of 0.88 °C in the four days) in operation, and (3) the high-accuracy virtual sensor with an average RMSE of 0.30 °C in a month (in February) from the in-situ calibration. The in-situ calibration in Case 3 decreased the initial virtualized sensor errors by 44.0 % and 32.9 %, respectively, for bias and random errors, by appropriately defining the correction function, the input feature, and the optimal calibration reference value in the mathematical calibration formulation. Based on the findings, two strategies for virtualization and calibration are proposed with their design tools for suitable problem formulation at the design stage, better accuracy in operation, and automated virtualization in the future.