Uncertainty calibration of building energy models by combining approximate Bayesian computation and machine learning algorithms

Bayesian analysis has attracted more attention in calibrating building energy models since it can naturally account for uncertainty of input parameters. However, standard Bayesian analysis requires to compute likelihood functions to represent the probability of the observed data under a statistical energy model. This makes the implementation of standard Bayesian analysis difficult for calibrating dynamic building energy models. Therefore, this research proposes a new Bayesian method by combining the approximate Bayesian computation with machine learning techniques in calibrating building models without computing likelihood functions. The results indicate that this new Bayesian computation approach combined with the machine learning algorithms can provide fast and reliable calibration for building energy models created using the EnergyPlus program. The calibration metrics in terms of the coefficient variations of root mean square errors from this new Bayesian method are well below the threshold value (15%) recommended from the ASHRAE standard for monthly energy data. Moreover, the better calibration results can be obtained by using linear or nonlinear post-adjustment techniques to account for the differences between the simulated and observed energy data. The method proposed here can significantly extend the application of Bayesian analysis in calibrating building energy models under uncertainty due to its simplicity and fast computation. Furthermore, the new Bayesian approach combined with machine learning technique can also be used to calibrate the simulation-based models for various types of energy systems.