The effects of multidimensional data clustering on the accuracy of virtual in-situ calibration in the photovoltaic/Thermal heat pump system

Abstract

Sensor errors have a significant negative impact on the control, diagnosis, and optimisation of building energy systems. The virtual in-situ calibration (VIC) method based on Bayesian inference and the Monte Carlo Markov chain can diagnose and calibrate sensor faults without installing new sensors. In previous studies, the historical data incorporated into the VIC method were typically measured under a certain steady-state operating condition without considering the multiple operating conditions of the system. In an actual photovoltaic/thermal heat pump (PV/T-HP) system, the overall system consists of a variety of operating conditions owing to changes in solar radiation, wind speed, or temperature. A sensor under different operating conditions produces different systematic and random errors, which affect the calibration. To solve this problem, the Gaussian mixture model (GMM) based on probability is applied to pre-process the data of various operating conditions, so that each steady-state operating condition has the same Gaussian distribution to ensure the consistency of errors. To verify the validity of the GMM clustering and the accuracy of the VIC method, four operating conditions are set according to the actual working state. The results indicate that the GMM can be used to cluster these mixed working conditions well, which guarantees the selection of the following steady-state measurements. In addition, the steady-state measurements of different operating conditions are incorporated into the VIC method, the systematic and random errors are reduced by 86% and 42%, respectively. Meanwhile, the systematic error accuracy for large faults increases by 12% (from 86% to 98%) compared with that for small faults after calibration, and the random error accuracy for large faults increases by 38% (from 42% to 80%) compared with that for small faults. In general, the VIC method based on GMM broadens the scope of sensor fault repair and lays a foundation for future applications in actual systems.