Parameter estimation for building energy models using GRcGAN

Parameter estimation methods can be classified into (1) manual (trial-and-error), (2) numerical optimization (optimization, sampling), (3) Bayesian inference (Bayes filter, Bayesian calibration), and (4) machine learning (generative model). Bayesian calibration has been widely used because it can capture stochastic nature of uncertain parameters. However, the results of Bayesian calibration could be biased by (1) the prior distribution assumed by the expert's subjective judgment; (2) the likelihood function that cannot always describe the true likelihood; and (3) the posterior distribution approximation method, such as the Markov Chain Monte Carlo, which requires significant computation time. To overcome this, a new approach using a generator-regularized continuous conditional generative adversarial network (GRcGAN) is presented in this paper. Five target parameters of the DOE reference building model were selected. GRcGAN was trained to estimate uncertain parameters using simulated monthly electricity and gas use. GRcGAN can successfully estimate five uncertain parameters based on 1,000 training data points. The proposed approach presents a potential for stochastic parameter estimation.

Automated summary

Parameter estimation methods can be classified into manual, numerical optimization, Bayesian inference, and machine learning. Bayesian calibration is widely used but can be biased by prior distribution, imperfect likelihood function, and computation time for posterior distribution approximation. This paper presents a new approach using GRcGAN to estimate uncertain parameters.