Dynamic graph-transformer architecture for multi-site photovoltaic power forecasting incorporating cloud cover dynamics and seasonal robustness in smart grid systems

Precise photovoltaic (PV) power prediction is essential to the stability of the smart grid but the quick irradiance changes due to clouds and dynamic spatial-temporal interactions make it difficult to predict accurately. Existing photovoltaic forecasting approaches exhibit limitations in jointly modeling dynamic spatial-temporal dependencies, cloud-induced irradiance variability, and seasonal distribution shifts, reducing forecasting robustness under real-world smart grid conditions. A novel DySAT-InfGraphFormer framework integrating dynamic graph construction, dual structural-temporal self-attention, graph-aware transformer encoding, and cloud-aware feature fusion is proposed for robust multi-site photovoltaic power forecasting. Python with PyTorch and PyTorch Geometric are used to implement the model and it was trained and tested using the UNISOLAR photovoltaic dataset. Implementation involves dynamic adjacency learning, feature normalization, dual-attention embedding, transformer refinement, and a regression forecast layer in an end-to-end architecture. Experimental findings show that the RMSE is reduced by 17.8 %, the MAE is reduced by 15.3 % and the accuracy of R 2 is increased by 9.6 % over state of the art baselines and seasonally stable. The given architecture is interesting as it merges the graph evolution and transformer intelligence into the cloud-conscious forecasting paradigm. The results verify that a combination of spatio-temporal graph attention and transformer encoding is much more effective in forecasting accuracy and resilience to real-life smart grid energy management.