Physics-Informed Predictive Control for a Fixed-Wing UAS under Uncertain Flight Dynamics

The variation in unmanned aerial system dynamics during flight poses significant challenges for predictive flight control when the presence of external disturbances and accurate aerodynamic models are not available throughout the flight envelope. This research proposes a physics-informed neural predictive control framework for the longitudinal dynamics of a fixed-wing unmanned aerial system. A neural predictor is trained using data collected at a limited set of flight points, while partial knowledge of the aircraft dynamics is incorporated through physics-based constraints during training. The resulting model is then used as the internal predictor of the model predictive controller and evaluated at operating conditions excluded from the training dataset. To account for residual model mismatch and external perturbations, a generalized proportional-integral observer is introduced to reconstruct a lumped disturbance. The estimated disturbance augments the predictive model rather than being assumed to act through the control channel, enabling the controller to account for disturbance effects that may not be directly cancelable. Numerical experiments using the UAS-S45 Bálaam system evaluate longitudinal reference tracking at trained and completely unknown operating conditions, with additional disturbance scenarios introduced exclusively during validation. The proposed framework combines physics-guided learning, online disturbance reconstruction, and predictive control to reduce dependence on an accurate global aircraft model while maintaining closed-loop performance under uncertain flight dynamics as evidenced by yielded results.