System-level virtual sensing method in building energy systems using autoencoder: Under the limited sensors and operational datasets

Abstract

Sensing networks and their environments are essential in intelligent building systems because of their increasing dependency on operational data. Virtual sensing technology has been applied in building energy systems to provide the more reliable and informative sensing environments. However, conventional virtual sensors still have structural and practical limitations under the physical sensor absences and limited datasets. Existing virtual sensors are separately developed by modeling multiple input variables and a single target (Xs to Y), which is the variable-level virtual sensor (VLVS); therefore, these virtual sensors cannot benefit by either using their target variable (Y) or by considering other virtual sensors when developing the models. This can result in insufficient accuracy, particularly in the limited sensors. Herein, to overcome these limitations, a novel virtual sensing framework, system-level virtual sensing (SLVS), is proposed for building energy systems using an autoencoder. Two strategies are also proposed. The autoencoder-based SLVS with the two strategies was applied in a real operational district heating system. The first strategy showed an improved accuracy using a new assistance virtual sensor, which is derived by additional information and knowledge regarding system design, control, and devices. It could also overcome the training data dependency in the limited datasets. The second strategy provided a replacement function for the SLVS specialized for backup and a calibration effect for the existing VLVS. Thus, the results showed that the suggested SLVS can achieve multifunctional high-accuracy virtual sensing; the accuracies of 99.89%, 99.68%, and 97.91% were shown respectively for temperatures, pressures, and control signals.