In-situ backup virtual sensor application in building automation systems toward virtual sensing-enabled digital twins

A backup virtual sensor (BVS) is widely used as a benchmark, for redundancy, and as a replacement for a physical sensor to implement intelligent operational applications in digital twin-enabled building systems. BVSs are modeled using various approaches and are employed to enhance physical sensor networks. BVSs are conventionally developed with built-in methods, that is, the model is built in a controlled environment, such as an experimental or simulation-based system. However, the built-in BVSs may show performance degradations when applied in real complex systems. This is because the real systems may contain unidentified practical uncertainties. Therefore, BVSs must be developed and managed with in-situ methods, that is, under real system environments. Nevertheless, the performance of an in-situ BVS is largely influenced by the training data and modeling methods. Thus, these factors must be considered in the BVS development and engineering process. Against this backdrop, this study suggests knowledge and industrial guidelines for developing high-performance BVSs in digital twin-enabled systems. Case studies are conducted for 64 cases (different training and testing datasets for 112 days) with a real building system to discuss BVS performance in terms of the training data periods, characteristics of the training data, and modeling methods. The research findings (1) offer modeling knowledge, industrial guidelines, a decision-making algorithm for appropriate modeling approaches, and insights into how building engineers implement virtual sensors and apply them to real building systems, and (2) demonstrate real-world applications of the virtual sensor-enabled digital twin in building automation systems.