To address the limitation in wind power forecasting accuracy caused by the difficulty of existing data-driven models in fully utilizing physical knowledge such as aerodynamic coupling in wind farms, a Physics-Informed Wind-aware Graph Attention Network (PI-WaGAT) is proposed. The model employs the Gaussian wake model to quantitatively analyze aerodynamic influences between turbines, constructing a dynamic topology that varies with wind conditions. On this basis, a wind-aware hypernetwork mechanism is designed to dynamically generate graph attention weights. By combining a Mixture Density Network with a physical feasible region regularization loss function based on conditional distributions, physical knowledge is adaptively embedded into the model training process. Experiments show that the proposed model outperforms benchmark models in both point and probabilistic forecasting accuracy, demonstrating good physical consistency under complex wind conditions.