In-situ sensor calibration for building HVAC systems with limited information using general regression improved Bayesian inference

Abstract

Sensors in building heating, ventilation and air-conditioning systems (HVACs) play important roles in maintaining indoor environmental quality and energy consumption. Owing to the repeatedly varied outdoor working environment and indoor users’ demand, sensor faults could be inevitable in the lifespan. To allow HVACs worked at fault-tolerant way, previous studies developed the in-situ sensor calibration method via energy conservation equations and Bayesian inference (EC-BI). However, the practical application may encounter challenges like limited-variable information, low-quality data and increasing risks of calibration uncertainty by indirect information supplement. These cause increasing in-situ calibration complexity and modeling costs. To address these challenges, this study proposed a general regression improved Bayesian inference (BI) in-situ sensor calibration strategy. The multiple linear regression (MLR) was utilized as a typical example of general regression method to improve the BI method. The proposed MLR-BI method was validated using both simulated and practical data of two building HVAC systems in two case studies. The principle component analysis (PCA)-based sensor fault reconstruction method was used for comparison under five fault conditions covering both single and simultaneous faults. Five variable scenarios were considered to validate the effectiveness of MLR-BI on HVACs with the limited variable information. Results indicated that the calibration accuracy of MLR-BI is over 99% under four conditions of the simulated case 1, which is about 6% and 8% higher than PCA and EC-BI respectively. For all the three variable scenarios of the simulated case 1, the calibration accuracy of MLR-BI is 99.65% on average. Especially in the four-variable scenario with limited variable information, MLR-BI shows the average calibration accuracy of 99.75% while PCA obtains 79.46% and EC-BI fails to work because of variable limitation. For the fault condition of the limitted-variable practical case 2, MLR-BI still outperforms the other two and obtains 97.1% calibration accuracy in two practical scenarios.