Comprehensive evaluation of the influence of meta-models on Bayesian calibration

Many efforts are used to eliminate the gaps between physical measurements and simulation results. The Bayesian calibration technique is one of the popular automatic calibration methods Allowing for all sources of uncertainty, the method attempts to correct inadequate models and parameters that produce discrepancies between prediction and observed data. Bayesian calibration requires numerous iterations of simulation. To reduce the computational cost, a meta-model is often used to replace the original detailed model. This paper evaluates and analyzes the influence of meta-model accuracy on the outcomes of the Bayesian calibration for building energy simulation model. The following five meta-models: multiple linear regression model, neural network, support vector machine, multivariate adaptive regression splines, and Gaussian process emulator are investigated and compared. The study varies the ranges of input parameters to explore the associated accuracy of the built meta-models. The calibration results are evaluated and compared using three criteria: simulation time, the coefficient of variation with root mean square error (CVRMSE) to true input parameter values, and CVRMSE to observed outputs (monthly and annual energy use intensity). The paper confirms that there are significant differences in simulation time and accuracy for different types of meta-models. The most accurate meta-model tested is the Gaussian process emulator (GPE), while it requires the longest computing time. The multiple linear regression model is the fastest but shows the worst performance in the calibration for the true input parameters. All the calibrated models, however, can predict well the outputs against the observations.