Zero-label cross-region solar radiation forecasting using domain adaptation-based transfer learning

초록

Accurate solar radiation prediction is essential for integrating photovoltaic (PV) systems into power systems and supporting supply–demand management. However, many regions lack on-site irradiance measurements due to limited sensor deployment. To address this, this study investigates domain adaptation (DA)-based transfer learning for solar radiation prediction under zero-label conditions, where solar radiation measurements are unavailable in the target region. Three representative DA methods: Domain Adversarial Neural Network (DANN); Deep Correlation Alignment (Deep CORAL); and Deep Adaptation Network (DAN) are systematically evaluated to examine how their relative effectiveness varies across cross-regional transfer conditions through 30 source-to-target transfer cases among six South Korean cities. The results show that DA models consistently outperform the non-adapted model by mitigating inter-regional discrepancies. The findings reveal that DA performance depends on both the magnitude and direction of inter-domain irradiance differences. Deep CORAL performs best in reducing overestimation in High-to-Low irradiance transfers, while DANN is more effective in mitigating underestimation in Low-to-High irradiance cases. Furthermore, embedding and feature importance analyses suggest that the DA methods adjust latent feature representations in distinct ways. This study demonstrates that reliable solar radiation prediction can be achieved even in regions without on-site solar radiation measurements and provides insights into how the relative effectiveness of different DA methods varies across cross-regional transfer conditions. Based on these findings, this study provides practical guidelines for selecting appropriate DA strategies, thereby supporting the deployment of prediction models in power systems where site-specific solar radiation measurements are unavailable.

키워드