🤖 AI Summary
To address the limited prediction accuracy in wind power forecasting caused by insufficient modeling of cross-sectional (inter-turbine) and temporal hierarchical dependencies, this paper proposes a multi-level coupled forecasting framework. Methodologically, it introduces, for the first time, a joint decoupling and co-optimization mechanism for three heterogeneous time-series dynamics—meteorological, turbine-level, and grid-level—and designs a hybrid deep learning model integrating graph neural networks, hierarchical attention, and physics-informed embeddings to explicitly encode spatiotemporal dependencies and physical constraints. Experiments across six real-world wind farms demonstrate an average 21.4% reduction in MAE and sub-3.2% error for one-hour-ahead forecasts, significantly outperforming state-of-the-art baselines. The core contributions are: (i) a novel hierarchical joint optimization paradigm enabling dynamic, cross-scale collaborative modeling; and (ii) empirical validation of physics-guided heterogeneous time-series fusion for enhanced forecasting reliability and interpretability.