Artificial neural networks for low-thrust spacecraft control

In this work, two artificial neural networks aimed at low-thrust spacecraft control are designed. One of the networks is called the controlling network; the another is called the predictive network. The controlling network is learned to solve the optimal control problem between two phase vectors. The predictive network is learned to propagate solutions of the extended equations of motion (prime and dual variables of the Pontryagin's maximum principle). The proposed networks can be used for orbital correction near a nominal regime. The architecture of the networks as well as sample construction process and learning methods are described. Simulation results are provided.