그린리모델링 활성화를 위한 기계학습 기반 EUI 예측 알고리즘 성능 비교 연구

초록

As carbon neutrality policies gain prominence, green remodeling has become increasingly important for improving the energy performance of aging public buildings. However, ECO2-OD-based energy performance evaluation, which is widely used in practice, requires numerous input variables, thereby limiting the rapid comparison of alternatives in the early decision-making stage due to the burden of data collection and the complexity of the input process. This study used real-world data from public facilities for the elderly and children collected through the 2020–2021 Regional Hub Platform Project for Green Remodeling of Public Buildings to compare machine learning-based surrogate models for approximating annual energy use intensity (EUI). Linear models, tree-based models, and deep learning models specialized for tabular data were evaluated under identical conditions. The results showed that tree-based models generally outperformed the others, and CatBoost was selected as the representative algorithm in terms of predictive accuracy and stability after hyperparameter optimization and multi-seed evaluation. In addition, a supplementary analysis of pre- and post-remodeling EUI changes showed that the selected model generally captured the direction of energy performance changes consistently. These findings provide practical criteria for selecting algorithms for surrogate model construction based on real project data and serve as a foundation for the future development of lightweight surrogate models with reduced input requirements.

키워드