Refining satellite trajectories with celestial body features using neural networks

Abstract

Satellite orbit propagation involves predicting the future position and velocity of a satellite from an initial position and velocity. Traditional physical models utilize information about the satellite and its environment to make such predictions. The SGDP4 model is widely used in this context due to its simplifications of the forces that act on the satellite, which enable high prediction speed albeit at the cost of a decrease in accuracy. This has motivated some authors to use machine learning techniques to refine predictions and reduce error. Existing proposals lack a thorough study of the effect of some decisions regarding the design of these models, and most importantly, do not incorporate features related to the state of celestial bodies that might have an influence on the satellite trajectory, thus remaining an unexplored field. In this paper, a novel model is proposed, incorporating such features at the initial time as well as through the entire prediction interval. This model in based on a neural network architecture that includes GRU layers, which effectively leverage sequential data. Our results show that the use of such sequential features greatly reduces the model error. Together with other design elements, such as the creation of independent models for specific coordinates and time windows, they provide a significant improvement in SGDP4 predictions in both the short and long term.