Paper Conference

Proceedings of Building Simulation 2017: 15th Conference of IBPSA


Comparison of Three Random Forest Models of a Chiller System

Han Sol Shin1, Cheol Soo Park1
1School of Civil, Architectural Engineering and Landscape Architecture, Sungkyunkwan University, Suwon, South Korea *Corresponding author

Abstract: The current building operation and maintenance is dependent on subjective decisions, e.g. building operator’s experience and knowledge, rather than employing a simulation model-assisted operation. It demands in-depth knowledge of building physics, systems and controls to develop the simulation model for optimal operation. Rather than using the first-principles based simulation tools, this paper presents a machine learning simulation model of a chiller in an office building. For this study, the BEMS data (a chiller’s electric energy, chiller supply water temperature, AHU return water temperature, AHU water flow rate, etc.) were collected from the existing office building (a total floor area: 21,577m2). The authors used a Random Forest (RF) method, one of the machine learning techniques. Three RF models were developed and cross-compared in this study as follows: Model A developed with 12 variables from BEMS data, Model B developed with the 12 variables plus 18 new variables constructed by two arithmetic operators (a total of 20 variables), Model C with the 12 variables plus 6 new variables constructed based on physics-based equations (a total of 18 variables). The CVRMSE of the three models are 8.56%, 5.44% and 4.28%, respectively.
Pages: 1631 - 1636