A Deep Learning-Based Bayesian Framework for High-Resolution Calibration of Building Energy Models

Abstract

Calibrating building energy models (BEMs) in high resolution is of significance to close the discrepancy between building modeling and field measurements. This is fundamental to support a diverse spectrum of applications for building sustainability and resilience analysis. However, with the Bayesian calibration as one of the most commonly used calibration methods, current calibration is mostly performed in low resolution (e.g., annual, or monthly), instead of high resolution, e.g., hourly or sub-hourly, which becomes increasingly crucial for rising BEM applications, such as demand response, building-renewable energy integration, and smart control. This is mainly attributable to the gaps in the existing Bayesian calibration process, including (1) over-parameterization and multi-solution in high-resolution calibration, (2) inadequacy of meta-model to capture high resolution building dynamics for calibration support, and (3) excessive computational burdens in likelihood function evaluation. Therefore, to close these gaps, this research proposes a novel deep learning-based Bayesian calibration framework, involving pre-calibration mechanism, Long Short-Term Memory as a surrogate model, and simplified covariance matrix calculation, to calibrate BEMs in high temporal resolution (i.e., hourly) with enhanced accuracy and computational efficiency. The application of this calibration framework in a case study demonstrates its effectiveness to calibrate a high-fidelity physics-based building model to simultaneously match measurements of heating, cooling, and electricity consumptions with the coefficient of variation of the root mean squared error (CV-RMSE) of <30% and the normalized mean biased error (NMBE) of <6% in the hourly resolution. Furthermore, the computation time for calibration significantly reduces by more than 99% (from an estimate of approximately 615.14 hours in the original Bayesian-based approach to ~1.67 hours for our proposed approach).

Keywords: Building energy modeling, Model calibration, Bayesian calibration, Deep Learning, Gaussian processes