In-Flight Hyperparameter Controller Adjustment for Load Alleviation of Very Flexible Aircraft

Model Predictive Control (MPC) has been previously proposed to enforce load constraints for very flexible aircraft. However, its practical use is limited by the large and time-varying computational cost of solving constrained optimizations in real time. This optimization problem can be particularly large sized in the case of very flexible aircraft. Various approaches have been developed to reduce the computational footprint of MPC including shortening the prediction horizon, move blocking, and modifying the weights in the cost function. While effective in reducing computational footprint, these approaches may also degrade closed-loop performance. To address the trade-off between the computational footprint and performance in load alleviation for very flexible aircraft, this paper proposes a novel data-driven methodology for dynamically in-flight MPC hyperparameter adjustment with respect to the maximum allowable CPU time. In our approach, an offline-trained neural network is utilized to capture the nonlinear mapping between MPC weights, computational time, and tracking error. Then this neural-network-based surrogate model is used online to adapt the weights in the MPC cost function thereby maximizing tracking performance under the computational feasibility constraints. Through simulations on a very flexible aircraft model, this paper demonstrates that the proposed approach can ensure constraint satisfaction and the required tracking accuracy while maintaining the computational footprint within the prescribed limits. Thus, the proposed framework provides a systematic and data-driven alternative to conventional offline trial-and-error tuning and enhances flexibility in balancing control performance versus in-flight computational requirements.