| Title |
Development of a Machine Learning Model for a Chiller using Random Forest Algorithm and Data Pre-processing |
| Authors |
Shin, Han-Sol ; Park, Cheol-Soo |
| DOI |
https://doi.org/10.5659/JAIK_SC.2017.33.9.67 |
| Keywords |
Inverse Modeling ; Machine Learning ; Random Forest ; Variable Selection ; Variable Construction |
| Abstract |
It has been widely acknowledged that a machine learning model can be used as a surrogate to a first-principle based dynamic simulation model. The accuracy and computation efficiency of a machine learning model is dependent on a combination of input variables. The random forest algorithm, one of the machine learning algorithms, can calculate a variable importance that determines the influence of each input variable on the output of the model. In this study, the authors developed three random forest models of a chiller in an existing building as follows: (1) Model A consisting of 12 measured variables from BEMS data, (2) Model B consisting of 2 measured input variables plus 4 new variables constructed by random selection, and (3) Model C consisting of 4 measured input variables plus 2 new variables constructed based on a physics-based equation. The CVRMSE of the three models are 8.56%, 5.44%, and 4.28%, respectively. The findings of this study can be summarized threefold: (1) all three random forest models are good enough to describe the dynamics of the chiller system, (2) the random forest machine learning algorithm can be used to develop a simulation model of the system, and (3) an accurate model can be constructed either by the random selection or the physics-based equation, even when a few input variables are given. |