LSTM-based wind disturbance prediction and compensation for UAV trajectory tracking control under time-varying wind conditions

Abstract Unmanned aerial vehicle (UAV) trajectory tracking in complex low-altitude environments is strongly affected by time-varying wind disturbances, abrupt gusts, and model uncertainty. This study proposes a learning-enhanced nonlinear model predictive control (L-NMPC) framework that combines online disturbance estimation, multi-step wind-disturbance prediction, and adaptive feedforward compensation. A multi-timescale exponentially weighted moving average estimator is used to reconstruct three-dimensional translational disturbances, while an Attention-LSTM network predicts their future evolution. The current disturbance estimate and multi-step predictions are adaptively fused according to disturbance magnitude and prediction horizon, and the resulting disturbance sequence is explicitly incorporated into the NMPC prediction model for anticipatory compensation. Under strong-wind conditions, the proposed predictor achieves an R 2 of 0.9678 for disturbance-magnitude prediction, with a median relative error of 3.1%. In Monte Carlo simulations, the proposed controller achieves a trajectory-tracking RMSE of 0.0240 m under standard conditions and a 95th-percentile error of 0.0343 m under robustness tests, reducing RMSE by up to 22.38% compared with the baseline controllers. These results demonstrate that the proposed LSTM-NMPC improves trajectory-tracking accuracy and robustness under time-varying wind disturbances.