Improved calibration of building models using approximate Bayesian calibration and neural networks

Abstract

Deep energy retrofits of buildings are crucial to meeting climate targets and depend on calibrated energy models for investor confidence. Bayesian inference can improve the rigour in standard practice and improve confidence in calibrated energy models. Approximate Bayesian computation (ABC) methods using neural networks present an opportunity to calibrate energy models while inherently accounting for parameter uncertainty, and face less computational burden than the current standard process for Bayesian calibration. A case study for a large, complex building is presented to demonstrate the applicability of ABC and parameter sensitivity screening is found to result in over-confidence in the resulting inference by between 14% and 85%. Finally, the presentation of posterior distributions as independent distributions may be misleading, which can misattribute the true likelihood of parameters. Highlights Implementation of an Approximate Bayesian Computation method incorporating the Sequential Monte Carlo algorithm with a neural network surrogate model. A comparison of Bayesian inference with standard practice. An investigation of sensitivity screening for parameter selection on the inference results. Application to a complex multi-zone dynamic energy model of a large retail building.