A solar radiation data generation method for solar energy utilization scenarios: BIPV generation forecasting as a case study

Accurate photovoltaic (PV) forecasting is crucial for grid stability, yet data-driven models are often hindered by the "cold-start" problem of data scarcity at new installations. Current generative models that directly synthesize power data act as "black-box" solutions, lacking physical interpretability and generalizability. To address this, we propose StochRad-UAGAN, a novel gray-box GAN framework. Instead of generating power data, it synthesizes physically-consistent solar radiation scenarios by modeling the key stochastic driver: the cloud-induced daily attenuation coefficient ( ). Our U-Net Attention GAN (UAGAN) proves superior to baseline models in capturing complex time-series features. In a BIPV forecasting case study, using our framework to augment a limited dataset reduced the prediction RMSE of various models by 10.1% to 49.7%, with an optimal generated-to-real data ratio of 2:1 to 4:1. This validates StochRad-UAGAN as an effective, generalizable solution to the data scarcity problem in solar applications, bridging deep learning with physical principles.