Paper Conference

Proceedings of Building Simulation 2021: 17th Conference of IBPSA

     

Dimensionality reduction for calibration of dynamic building simulation models

Aurelien Bres
AIT Austrian Institute of Technology, Austria

DOI: https://doi.org/10.26868/25222708.2021.30778
Abstract: Calibrating dynamic building simulation with subhourly data involves comparing a high number of simulated and measured values. The predictability and informativeness of each value taken separately may be limited. Thus, dimensionality reduction, by transforming the simulation output data to a space of lower dimension while retaining significant features, has the potential to improve the computational efficiency of calibration. This contribution aims at investigating the use of linear and nonlinear dimensionality reduction techniques for the calibration of dynamic building simulation models involving uncertainties pertaining to both building and heating systems. Two calibration cases are considered. The techniques used for dimensionality reduction are principal component analysis on the one hand, and on the other hand nonlinear feature extraction based on artificial neural networks. A Bayesian calibration method is then applied to the outputs with reduced dimension, and its performance is assessed with synthetic data. The results show improved performance in comparison to calibration based on daily average values.
Keywords: dimensionality reduction, Bayesian calibration, multivariate outputs, calibration
Pages: 1991 - 1998
Paper:
bs2021_30778