基于图神经网络的四旋翼风扰观测与敏捷控制

Robust and agile flight under strong wind disturbances remains one of the primary objectives in disturbance-rejection control for unmanned aerial vehicles (UAVs). Conventional observer-based control approaches often fail to effectively model the influence of aggressive wind fields on UAV dynamics, leading to significant performance degradation in complex wind scenarios. This paper proposes a lightweight wind-effect predictor based on a graph neural network (GNN) with adaptive channel extraction, leveraging the differential flatness property of quadrotors. The predictor is trained using historical flight trajectories to enable real-time estimation of complex wind disturbances encountered during flight. A model predictive control (MPC) scheme is then developed to achieve disturbance-aware trajectory tracking. Specifically, flight data are collected under various wind conditions, including sustained winds, gusts, and crosswinds. A total of 995650 trajectory residual samples are used to train the predictor, and 150000 samples are reserved for testing and benchmarking against state-of-the-art models. Experimental results show that the proposed wind-effect predictor achieves a 14.9% improvement in prediction accuracy compared to existing methods. Moreover, the wind-aware MPC based on this predictor improves trajectory tracking performance by 50.35% under aggressive wind disturbances. Finally, real-world flight experiments on a custom-built quadrotor platform validate the effectiveness of the proposed method, demonstrating a 27.35% enhancement in trajectory tracking accuracy under wind disturbances.