다변수 기반 기계학습을 통한 초등학교 연간에너지 소비량 예측 모델 성능비교

초록

This study aims to identify suitable machine learning models for predicting the annual energy consumption of elementary school buildings and to examine explanatory variables that improve prediction accuracy. Energy consumption data from 2016 and 2017 were used for training, while 2018 data were used for validation. The models evaluated included Linear Regression, Neural Networks, Random Forest, and Gradient Boosting, with the number of explanatory variables ranging from 4 to 10. Among the models, the Random Forest model with 10 variables showed the best performance, achieving a MAPE of 22.4 percent and an R² value of 0.821. Annual water consumption was found to have a significant effect on improving prediction accuracy, particularly in buildings with high energy use. Although further improvements are needed before the model can be applied to policy decisions, it may serve as a useful reference for identifying unusual energy consumption patterns in school buildings. Future studies could improve the model by optimizing parameters, incorporating more detailed local climate data, and including variables that reflect the specific operational characteristics of schools. © 2026 Architectural Institute of Korea.

키워드