Quantifying the effects of different data streams on the calibration of building energy simulation

Bayesian calibration of building energy simulation (BES) has gained growing attention for its capability to tackle uncertainties and narrow the gap between simulated and measured results using expert knowledge. However, how output or parameter correlations influence single-output Bayesian calibration (SOBC) and multiple-output Bayesian calibration (MOBC) for BES has not been investigated and compared. Hence, this paper intends to determine the impacts of output or parameter correlations and data informativeness on BC of BES. Aiming to leverage multiple outputs' correlations while comparing with traditional SOBC in calibration performance and computation cost, we also developed the MOBC model. Compared to including weakly correlated outputs or parameters, the results show that strongly correlated outputs or parameters in MOBC reduce the Coefficient of Variation of the Root Mean Squared Error (CVRMSE) by 5.56% and 4.729% respectively and bring notably better Continuous Ranked Probability Score (CRPS) results. Conversely, including strongly correlated parameters in SOBC causes worse model performance. These results reflect that SOBC has parameter identifiability issues that MOBC can solve. The findings contribute to our better understanding of the impacts of (1) output or parameter correlations and (2) different data streams' informativeness on the calibration performance of SOBC and our developed MOBC for BES.

Automated summary

Bayesian calibration of building energy simulation (BES) using expert knowledge can reduce uncertainties and improve the agreement between simulated and measured results. This study investigates the effects of output and parameter correlations, as well as data informativeness on the calibration performance of single-output Bayesian calibration (SOBC) and multiple-output Bayesian calibration (MOBC) for BES.