| Title |
Epistemic and Aleatoric Uncertainty of Bayesian Neural Network Model for a Chiller |
| Authors |
김재민(Kim, Jae Min) ; 박철수(Park, Cheol Soo) |
| DOI |
https://doi.org/10.5659/JAIK.2020.36.6.177 |
| Keywords |
Bayesian Neural Network; Epistemic uncertainty; Aleatoric uncertainty; Outliers; Support Vector Data Description |
| Abstract |
Because the machine learning model is a black-box model, it is difficult to quantify the causality between inputs and outputs. In addition,
the model is influenced by its inherent uncertainty in describing a system’s behavior of interest. In order for a machine learning model to be
reliable, its prediction performance as well as uncertainty must be quantified together. Bayesian Neural Network (BNN) is a useful tool to
describe stochastic characteristics of deep learning models by estimating distributions of the models’ weights. Model uncertainty in BNN can
be classified into epistemic and aleatoric uncertainties. Epistemic uncertainty is caused by lack of data or knowledge. In contrast, aleatoric
uncertainty is caused by outliers or noises inherent in training data and can be reduced by removing abnormal data from the training dataset.
In this study, the BNN models were developed for a compression chiller in an existing office building with BEMS data, and then epistemic
and aleatoric uncertainties were analyzed. It is found that both uncertainties are significant in the simulation model even though the model’s
accuracy is satisfactory with the CVRMSE of less than 15%. It is suggested that before attempting to apply the machine learning model to
real applications, the both uncertainties must be carefully analyzed. It is recommended that the both uncertainties can be reduced by adding
more data as well as removing outliers. |