전이 학습 기반 한반도 전역 태양 복사 예측 모델 개발

Abstract

Previous studies applying transfer learning to solar irradiation prediction have been limited to specific regions. Therefore, this study aims to develop a solar radiation prediction model applicable across the Korean Peninsula by validating the effectiveness of transfer learning using a Bi-GRU combined with DANN across six regions in South Korea. Experimental results showed that the proposed model achieved high predictive performance in most regions, with Daejeon as the source domain yielding the best results (RMSE = 42.98, MAE = 30.36, R² = 0.9753). Furthermore, the prediction performance was comparable to that of conventional learning (source = target), showing no significant degradation. These findings suggest that a Daejeon-based model can be effectively applied for solar irradiation prediction throughout South Korea. Future research should consider incorporating seasonal characteristics of solar irradiation to further enhance model performance.