Adaptive Perceptual Graph Learning for Robust Wind Power Forecasting Under Incomplete Observations

Reliable wind power forecasting plays a crucial role in enabling efficient scheduling and effective energy management in virtual power plants. However, incomplete data observations caused by sensor malfunctions and volatile environmental conditions significantly compromise the reliability of wind power forecasting. To address these challenges, this paper presents a robust adaptive perceptual graph learning framework designed to enhance forecasting accuracy and robustness. The proposed framework uses an attention-gated GRU encoder to restore spatiotemporal dependencies under incomplete observations and integrates an adaptive graph convolution module to dynamically capture complex, non-stationary spatial correlations between wind turbines. To further promote physical consistency, a physics-informed discriminator is introduced as a plug-in component, guiding the model to align its outputs with the nonlinear aerodynamic relationship between wind speed and power generation. In addition, a sensitivity-aware adjustment strategy is developed to dynamically emphasize low-speed operating regimes. Comprehensive experiments conducted on real-world wind farm datasets show that the proposed method consistently surpasses baseline approaches, achieving a reduction of up to 35% in mean absolute percentage error, even under severe missing data scenarios.