Fault detection and calibration for building energy system using Bayesian inference and sparse autoencoder: A case study in photovoltaic thermal heat pump system

The rise of clean energy such as solar energy provides a new idea to optimize the energy structure, and Photovoltaic thermal (PVT) heat pump system is one of the mainstream development at present. The wrong sensor data will have a great impact on the control and efficiency of the whole PVT heat pump system. In order to deal with this situation, Virtual in-situ calibration (VIC) based on Bayesian inference and Markov chain Monte Carlo (MCMC) is applied to PVT system. This paper presents a generic sensor fault diagnosis and calibration method for building energy systems using a PVT heat pump system as an example. Firstly, the data generated based on the mathematical model of PVT system is used to test the feasibility of VIC in this system. The results show that VIC can well reduce the systematic error and random error of the sensor. However, with the further study, it is found that in the actual system, it is impossible to establish a mathematical model which can reflect the relationship between the sensors with high accuracy. The VIC may fail or even have worse calibration results. Therefore, in order to solve the above problems, this study suggests using sparse autoencoder model instead of mathematical model. The model proved to be more accurate and to reflect the interconnectedness of the sensors, which helped the implementation of the VIC. The calibration method based on sparse autoencoder (SAE) can calibrate not only a single sensor in the PVT heat pump system, but also multiple sensors in a local area. After sparse autoencoder Virtual in-situ calibration (SAE-VIC) calibration, systematic and random errors of all sensors can be effectively reduced, and the accuracy of sensor calibration can reach more than 90 %.