Journal of Korean Institute of Architectural Sustainable Environment and Building Systems ISSN:1976-6483(Print) 2586-0666(Online) 한국건축친환경설비학회논문집

Building energy benchmarking is a crucial step in building energy monitoring. Current benchmarking practices often face limitations rooted in the adopted methodologies. For instance, existing benchmark tools are based on the ordinary least squares regression method, which is highly sensitive to outliers and non-linearity in the dataset; this is likely to produce inaccurate benchmark scores. We propose a quantile regression benchmarking model reinforced with a gradient boosting decision tree algorithm (LightGBM). On the one hand, quantile regression analysis estimates the conditional distribution across quantiles, better capturing non-linearity in building energy behavior at various hierarchies of energy usage. On the other hand, the tree-based algorithm corrects previous errors during model training for better optimization. The said models were developed using a dataset from kindergarten school buildings in South Korea consisting of 3,090 buildings. From the developed models, it was observed that total floor area is a major determining factor in the electric energy use of kindergarten buildings in Korea, while the number of floors was also an energy driving factor in middle energy-consuming buildings. The proposed model provides accurate estimations of building energy performance, better facilitating energy comparison of a given building with its peers. The developed models provide valuable insights for energy conservation programs that might later inform building policies and for individuals wishing to reduce the energy consumption of their facilities.