Solar-Net: Adaptive Fusion of Spatial-Temporal Features for Resilient Solar Power Generation Forecasting

Solar power generation forecasting faces significant challenges due to intermittency and volatility, particularly under extreme weather conditions. This study proposes Solar-Net, a novel solar power generation prediction model based on a CNN+Transformer hybrid parallel architecture with an adaptive attention fusion mechanism. The CNN branch extracts spatial features from the power station layout and environmental conditions, while the Transformer branch models temporal dependencies in generation patterns. The core innovation lies in the adaptive attention fusion mechanism that dynamically adjusts branch weights according to real-time meteorological conditions, enabling the model to automatically adapt to varying environmental scenarios. Experiments were conducted on a comprehensive dataset containing over 50,000 observation points from two photovoltaic power stations. Results demonstrate that Solar-Net achieves superior performance compared to existing methods, with Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) improvements of 12.7% and 10.9%, respectively. Under extreme weather conditions such as dust storms, the model maintains prediction errors within 8.5% of peak power generation, representing a 45.7% average reduction compared to baseline methods. The multi-scale convolution design enhances prediction accuracy by 10.5% while reducing computational complexity by 21.3%. The proposed Solar-Net model provides a robust and efficient solution for solar power generation forecasting, demonstrating significant potential for improving grid dispatching efficiency and supporting renewable energy integration in power systems.

@article{Su2025,
  title     = {Solar-Net: Adaptive Fusion of Spatial-Temporal Features for Resilient Solar Power Generation Forecasting},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {6},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Wenqian Su and Jason See Toh Seong Kuan and Xiangyu Shi and Yuchen Zhang},
  doi       = {10.14569/IJACSA.2025.0160626},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160626}
}