Learning of Energy Primitives for Electrified Aircraft

With the increasing electric loads present onboard electric aircraft, electrical components now support safety-critical functions for which component failure may cause catastrophic consequences. Throughout a mission, safety must be considered for power and thermal constraints within the powertrain and for trajectory constraints in the vehicle dy-namics. While enforcing comprehensive safety guarantees may be challenging for developmental systems, learning a safe set and using a terminal constraint can be used to guide the system towards known safe behaviors. To provide modularity in the learning rather than learning only a single trajectory, the basic behaviors of the system can be divided into primitives and learning can be conducted for each primitive. While existing learning techniques using a safe set require knowledge of feasible safe trajectories a priori, this paper presents a learning technique which iteratively improves performance through learning an unknown safe set for energy primitives. This paper also presents a reachability technique to ensure safe transitions between primitives which exhibit different dynamics. This technique exploits the structure of the dynamics to minimize branching among hybrid modes. For a simulated case study, the learning model predictive controller results in vehicle safety at all time steps while matching the performance of an unsafe controller. Additionally, the controller demonstrates iterative improvements in performance for a mission where multiple primitives are used.