Two-stage weather-aware unlabeled sky image classification and solar irradiance forecasting based on residual attention network

Accurate solar irradiance forecasting is critical for the reliable integration and operation of photovoltaic (PV) systems, yet remains challenging due to strong atmospheric variability and the limited availability of scalable, consistent weather annotations. This paper proposes a two-stage, weather-aware image-based framework that jointly addresses sky-condition classification and solar irradiance forecasting using unlabeled ground-based sky images. In the first stage, a Residual Atmospheric Feature Encoder augmented with channel–spatial attention and physics-inspired convolutional regularization is employed to extract discriminative atmospheric representations. Unsupervised clustering combined with heuristic pseudo-labeling based on radiometric and cloud-related image statistics is then used to derive physically meaningful sky-condition labels, enabling supervised training without manual annotation. In the second stage, the inferred weather regimes are explicitly integrated into a regime-aware solar irradiance forecasting model, reducing regime-mixing effects and improving physical consistency in prediction. Extensive experiments conducted on long-term benchmark sky-image datasets demonstrate strong and stable classification performance, achieving 94.37% accuracy and 94.58% F1-score in the full multi-class setting, improving to 96.82% accuracy for irradiance-based three-class grouping and 98.08% accuracy for binary clear versus non-clear classification. Weather-stratified forecasting analysis reveals distinct regime-dependent error characteristics: clear-sky conditions yield the lowest forecasting error (RMSE = 28.61 W/m 2 , R 2 = 0.989), while partly and mostly cloudy conditions exhibit higher uncertainty due to intermittent cloud dynamics, and overcast and rainy regimes show more stable but attenuated irradiance behaviour. Applicability-domain analysis confirms that weather-aware conditioning constrains predictions to physically plausible operating regimes and mitigates the influence of anomalous samples.