Toward label-efficient and comprehensive solar irradiance estimation with SkyMAE self-supervised learning on all-sky images

Precise measurement of Global Horizontal Irradiance (GHI), Direct Normal Irradiance (DNI), and Diffuse Horizontal Irradiance (DHI) is essential for advanced solar energy applications. While traditional radiometric sensors offer high reliability, their prohibitive costs and maintenance requirements hinder large-scale deployment, thus driving a research shift toward all-sky image-based approaches. However, most existing methods employ a single-task framework and rely heavily on auxiliary sensor data, thereby failing to fully exploit the rich information contained in all-sky images. Furthermore, traditional supervised transfer learning strategies may exhibit reduced transferability when adapting knowledge to lightweight models or sites with heterogeneous climatic conditions and native data resolutions, making them more susceptible to architectural mismatches and negative transfer. In this paper, a SkyMAE framework is developed, in which a self-supervised Masked Autoencoder (MAE) pretrains on all-sky images to learn visual representations, and the learned representations are subsequently transferred to a unified multi-task estimation model to simultaneously generate real-time estimates of GHI, DNI, and DHI without reliance on auxiliary sensor data. Extensive evaluations on a customized lightweight model across climatically different target sites demonstrate that SkyMAE improves estimation accuracy and data efficiency. Using the full training datasets, Local SkyMAE initialization reduces RMSE compared with Random initialization on both target sites (Folsom: 26.4%, 8.3%, and 20.5%; NREL: 28.4%, 23.2%, and 30.1% for GHI, DNI, and DHI estimation, respectively). Moreover, Local SkyMAE pretraining outperforms cross-site supervised transfer on the target site (NREL). Further analysis reveals that pretraining on large-scale unlabeled external datasets can help mitigate the scarcity of local training data, while enhancing the relevance of pretraining imagery to the local domain can further contribute to improved irradiance estimation performance. These findings suggest that the proposed SkyMAE framework provides a promising image-based approach for solar resource assessment across different locations and climatic conditions.