Design and Optimization of a Resident Space Object's Orbit Based on Machine Learning

Orbit prediction accuracy remains critical for space situational awareness, yet traditional physics-based models (e.g., SGP4) face limitations due to incomplete space environment data, noisy trajectory measurements, and undisclosed satellite maneuvers. Recent studies highlight persistent errors in Two-Line Element (TLE) datasets and inadequate modeling of periodic orbital dynamics using polynomial approximations, underscoring the need for novel approaches. This research addresses the challenge of improving orbit prediction accuracy under incomplete information by developing a hybrid methodology that synergizes machine learning with classical curve fitting. The proposed approach integrates Long Short-Term Memory (LSTM) neural networks with adaptive curve-fitting techniques. By training on historical TLE datasets spanning multiple orbital regimes, the LSTM architecture learns temporal dependencies and nonlinear patterns, while curve fitting captures deterministic orbital mechanics.Implementation demonstrates a 23-41% reduction in positional errors compared to SGP4 predictions across low-Earth and geostationary test cases. The methodology successfully mitigates error accumulation from TLE noise and unmodeled perturbations, validated through 18-month orbital propagation tests. Case studies of collision-prone satellites confirm improved maneuver detection sensitivity and 72-hour prediction reliability. The hybrid framework outperforms conventional physics-based models under realistic data constraints, offering enhanced collision avoidance capabilities. Future work will focus on real-time model updating and integration with space surveillance networks to support sustainable space operations.