Model Predictive Control of Quadcopter Using Physics-guided Neural Network

Nonlinear models are being utilized in MPC frameworks for UAVs because they are able to enhance control performance by precisely representing the physical system and generating more accurate prediction horizon. However, uncertainty in parameters and approximation errors associated with physics-based (PB) models could dramatically degrade MPC performance. On the other hand, data-driven models require a large amount of training data with salient representation throughout the operational range, which, however, is often difficult to obtain. To address these limitations, this research presents a new physics-guided neural network (PGN) model that adopts the RNN structure as its backbone and embeds the residuals computed by the PB models to enforce physical constraints. Thus, the proposed PGN can be trained with a smaller amount of data compared to the purely data-driven networks, and even precisely represent the system dynamics beyond the range of the training data. Numerical case study is performed to construct a PGN model to represent the quadcopter UAV, which is then employed in MPC for trajectory tracking. The PGN-MPC demonstrates better performance (tracking accuracy and robustness) in comparison to MPC based on the nominal PB model and fully-connected neural network (FCN) model.