🤖 AI Summary
To address the insufficient utilization of meteorological information and the trade-off between accuracy and efficiency in short-term renewable energy generation forecasting, this paper proposes a deterministic forecasting method integrating spatiotemporal coupled attention mechanisms with a lightweight physics-constrained network. It is the first to explicitly embed heterogeneous, multi-source meteorological observations into the forecasting framework, enabling joint modeling of multi-site meteorological data. The method synergistically combines graph neural networks, numerical weather prediction downscaling, and physics-informed regularization to enforce physical consistency. Evaluated on real-world wind and photovoltaic power plants, the approach achieves a 21.3% reduction in mean absolute error and a 3.8× speedup in inference latency, significantly enhancing both forecasting accuracy and real-time performance—thereby meeting stringent grid dispatch requirements.