Calibration study of uncertainty parameters for nearly-zero energy buildings based on a novel approximate Bayesian approach

Buildings face various uncertainties in operation, leading to significant discrepancies between actual energy consumption and design-phase predictions. Thus, building energy calibration becomes critically important. Bayesian calibration is widely used for uncertainty in calculations; however, its application is limited in complex models where the likelihood function cannot be expressed analytically form or its computation is prohibitively expensive. To address this, this study introduces a novel Approximate Bayesian Calibration (ABC) method based on a Particle Filter (ABC-PRC), which employs a two-step approximation to avoid the direct computation of the likelihood function. The calibration process began with the identification of key uncertain parameters in a nearly-zero energy building (NZEB) through global sensitivity analysis. Subsequently, a machine learning model was trained as a surrogate model. Finally, the uncertain parameters were calibrated using the ABC-PRC method, and the results were compared with the Population Monte Carlo (ABC-PMC) method and the Sequential Monte Carlo (ABC-SMC) method in the ABC method. The results indicate that the proposed ABC-PRC method demonstrates significant advantages in both accuracy and stability. The 95 % confidence interval of the CvRMSE consistently remains below 6 %, while the 95 % confidence interval of the NMBE stays within the range of −2 %–3.8 %. These values are significantly lower than the requirements of 15 % and ±5 % specified in relevant standards. Therefore, this study presents a novel and effective method for building energy consumption calibration, offering significant practical value.