Review on building energy model calibration by Bayesian inference

A building energy model (BEM) is essential for understanding building energy consumption, evaluating energysaving measures, and developing associated codes, standards, and policies. The calibration of BEM helps to ensure the accuracy of the model, whereas it remains a challenge. Conventional manual or automated methods are mostly deterministic and neglect the inherent uncertainties of BEM. In comparison, the recent development of the stochastic BEM calibration based on Bayesian inference has gained attention, whereas many are baffled by its underlying theory, strengths, limitations, and implementations. There are also various mathematical models and tools in the literature, making it hard for selection. This paper aims to unravel the myths about the Bayesian inference and critically review various implementation options with a series of model selections suggested so that a user would be able to employ the Bayesian inference calibration at the end of the paper. We also hope that the review contributes to facilitating a broader implementation of the method for BEM calibrations. First, an overview is summarized for the current status and development of Bayesian inference calibration in building energy modeling. Second, the theory and methodology of model calibration, Bayesian statistics, and Markov Chain Monte Carlo are illustrated. Third, the implementation of Bayesian inference is described, including several practical issues such as BEM determination, unknown calibration parameters number, their ranges and distributions, Meta-model selections, and programming languages based on the statistical package R. The review ends with conclusions and future work identified.

Automated summary

Building Energy Model (BEM) calibration is crucial for accuracy, with recent focus on stochastic Bayesian inference calibration. However, confusion remains regarding theory, strengths, limitations, and implementations. Selecting appropriate mathematical models and tools poses a challenge.