Predicting Future Orbital States Across Three Different Orbit Types Using Artificial Neural Networks
Nayana Jayakumar
09/09/2026
With the rapid growth in space exploration and commercialization, the number of Resident Space Objects (RSO’s) has significantly increased, raising the risk of orbital collision. As of September 2025, there are around 23030 satellites in Earth’s orbit with more than 650 estimated collisions, according to the European Space Agency.
This data highlights the need for improved methods to accurately predict satellite trajectories and mitigate the collision risk. However, classical propagators, like SGP4, suffer from uncertainty due to perturbations. Most of the Machine Learning (ML) research done in this area explores satellites and debris orbiting the earth in Low Earth Orbits (LEO), Geosynchronous Earth Orbits (GEO) or both, with varying success . These researches were done using either Recurrent Neural Networks (RNNs) like Long Short Term Memory (LSTM) or Reinforcement Learning (RL) models.
This project proposes Astraeus, a set of six Artificial Neural Network (ANN) models that can generalize across LEO, Medium Earth Orbits (MEO), and GEO; based on an input of the Keplerian elements. Using Two-Line Element (TLE) data from Space-Track, Keplerian elements were converted into position and velocity state vectors using SGP4 for training. Astraeus learns from historical orbital data to predict future satellite state vectors without relying on analytical approximations. Results show that Astraeus achieved an overall prediction accuracy of 94%, with position vectors exhibiting lower percent error than velocity components. These findings demonstrate the potential of machine learning to enhance satellite navigation and space traffic management systems by improving orbit prediction across multiple orbits.