Attention Convolutional Neural Network Model Estimation of Solar Energy Potential Using Radiance Sky Imagery

The rapid growth of renewable energy adoption has heightened the need for accurate solar energy prediction to ensure grid stability, particularly in regions with high solar penetration. However, traditional forecasting methods relying on historical meteorological data often fail to address short term fluctuations caused by dynamic cloud movements, limiting real time adaptability. To overcome this challenge, this study proposes a deep learning framework integrating convolutional neural networks (CNNs) with attention mechanisms to predict photovoltaic (PV) output from radiance sky images. Two datasets capturing diverse sky conditions were used to evaluate three architectures: a baseline CNN, CNN with Squeeze-and-Excitation (SE) Attention, and CNN with Spatial Attention. The CNN with SE-Attention model significantly outperformed baseline models, reducing prediction errors and improving explanatory power, as validated by metrics including RMSE, MAE, and R². Gradient-weighted Class Activation Mapping (GradCAM) further demonstrated the model’s ability to prioritize meteorologically critical regions, such as cloud edges and solar disk areas, with distinct attention patterns for sunny and cloudy scenarios. The framework’s practical utility was enhanced through deployment in an interactive web-based Graphical User Interface, enabling real-time solar potential simulations for energy operators. By combining attention mechanisms with interpretable design, this work advances short-term solar forecasting accuracy while providing actionable insights for grid management. Future research directions include multi-modal data fusion and hybrid transformer-CNN architectures to improve robustness across diverse climatic conditions.