실내 환경 예측 모델의 파인튜닝을 통한 대상지별 적응 기간 비교 = Comparison of Adaptation Periods by Sites through Fine-Tuning of Indoor Environment Prediction Models

다국어 초록 (Multilingual Abstract)

Purpose: This study aimed to quantify how quickly indoor environmental prediction models adapt when transferred between different buildings using fine-tuning–based transfer learning. By focusing on temperature, CO2, and PM2.5 prediction, the study examines both the initial performance degradation caused by domain shift and the subsequent recovery process, in order to identify a practical adaptation period for reliable model deployment.
Method: DNN models trained on long-term data from an office building (Site A) were transferred to two large commercial buildings (Sites B and C). For each target site, daily fine-tuning was performed for 7 days using sliding-window datasets from the previous week, and day-by-day MAE, CvRMSE, and R2 were evaluated. Result: n Site B, temperature and CO2 models showed severe errors in the first 1~2 days (CvRMSE up to 91.51% and 60.78%) but converged to stable levels after about 3–5 days of fine-tuning, whereas PM2.5 remained more variable (CvRMSE about 9~28%). The results indicate that several days of fine-tuning are required for reliable model transfer between buildings, and that PM2.5 prediction in particular needs longer training and further model refinement.