A Generative Sky Image-Based Two-Stage Framework for Probabilistic Photovoltaic Power Forecasting

Solar forecasting using ground-based sky image offers a promising approach to reduce uncertainty in photovoltaic (PV) power generation. However, existing methods often rely on deterministic predictions that lack diversity, making it difficult to capture the inherently stochastic nature of cloud movement. To address this limitation, we propose a new two-stage probabilistic forecasting framework. In the first stage, we introduce I-GPT, a multiscale physics-constrained generative model for stochastic sky image prediction. Given a sequence of past sky images, I-GPT uses a Transformer-based VQ-VAE. It also incorporates multi-scale physics-informed recurrent units (Multi-scale PhyCell) and dynamically weighted fuses physical and appearance features. This approach enables the generation of multiple plausible future sky images with realistic and coherent cloud motion. In the second stage, these predicted sky images are fed into an Image-to-Power U-Net (IP-U-Net) to produce 15-min-ahead probabilistic PV power forecasts. In experiments using our dataset, the proposed approach significantly outperforms deterministic, other stochastic, multimodal, and smart persistence baselines models, achieving a superior reliability–sharpness trade-off. It attains a Continuous Ranked Probability Score (CRPS) of 2.912 kW and a Winkler Score (WS) of 33.103 kW on the test set and CRPS of 2.073 kW and WS of 22.202 kW on the validation set. Translating to 35.9% and 42.78% improvement in predictive skill over the smart persistence model. Notably, our method excels during rapidly changing cloud-cover conditions. By enhancing both the accuracy and robustness of short-term PV forecasting, the framework provides tangible benefits for Virtual Power Plant (VPP) operation, supporting more reliable scheduling, grid stability, and risk-aware energy management.