Multivariate modeling on wake-affected wind farms by two-stage hybrid graph neural network

Accurate wake effect modeling is essential for wind farm optimization, as it can significantly improve wind energy utilization efficiency, thereby accelerating the achievement of ‘Dual Carbon’ strategic goals. However, existing approaches struggle to balance computational efficiency with physical accuracy. To address these challenges, a two-stage hybrid graph neural network (TSGNN) model for multivariate wind farm wake effect modeling is proposed. Firstly, an adaptive directed graph structure based on a physical cone model is designed, which dynamically captures the directionality and spatial extent of wake interactions under real-time wind conditions while eliminating redundant topological connections. Secondly, a two-stage hybrid graph neural network is developed, applying specialized processing based on the distinct roles of boundary versus internal turbines to accurately capture complex flow phenomena. Thirdly, an adaptive weight mechanism is proposed to automatically learn the optimal balance between geometric wind-direction relationships and aggregated neighborhood information, improving learning efficiency. Finally, the effectiveness of the proposed approach is validated by extensive experiments on both simulated and real-world data. Four critical parameters, including effective wind speed, turbulence intensity, and two structural loads, are successfully predicted simultaneously by the TSGNN. Compared to baseline models, the proposed method achieves the highest prediction accuracy while maintaining the spatial prediction consistency index above 92 %. Moreover, it achieves an order-of-magnitude higher computational efficiency than physics-based tools, enabling real-time applications. Overall, a practical solution for optimizing wind energy management and utilization in large-scale wind farms is offered by the proposed model.