Session

Technical Session 9: Advanced Technologies III

Location

Utah State University, Logan, UT

Abstract

The current industry standard orbital propagator, the Simplified General Perturbation Model-4 (SPG4), relies completely on physics-based orbital mechanics, can only provide accurate orbital predictions ~12 hours in advance. We developed a novel hybrid model, combining the SGP4 baseline with two machine learning estimators, autoencoder and random forest, in order to reduce the errors of the SGP4 propagator. The sources of errors in SGP4 propagators come from incomplete perturbation calculations and low-order of series expansions. The time-series nature of these error patterns are modeled by our machine learning estimators and then are used to make corrections to the SGP4 propagation, which result in more accurate orbit predictions. We tested our hybrid model on 3 satellite objects with the corresponding TLE (Two Line Element) data. The improvement on orbit prediction achieved 20-30% over the future 30 days period. The limitation of this hybrid approach is the requirement of at least 3 years of historical TLE data for the machine learning models, but could be overcome by creating synthetic orbital data from a similar space object. This hybrid model can be easily packaged into a software tool for space mission operation planning and facilitate mission autonomy.

Available for download on Saturday, August 07, 2021

Share

COinS
 
Aug 12th, 9:00 AM

Improved Orbital Propagator Integrated with SGP4 and Machine Learning

Utah State University, Logan, UT

The current industry standard orbital propagator, the Simplified General Perturbation Model-4 (SPG4), relies completely on physics-based orbital mechanics, can only provide accurate orbital predictions ~12 hours in advance. We developed a novel hybrid model, combining the SGP4 baseline with two machine learning estimators, autoencoder and random forest, in order to reduce the errors of the SGP4 propagator. The sources of errors in SGP4 propagators come from incomplete perturbation calculations and low-order of series expansions. The time-series nature of these error patterns are modeled by our machine learning estimators and then are used to make corrections to the SGP4 propagation, which result in more accurate orbit predictions. We tested our hybrid model on 3 satellite objects with the corresponding TLE (Two Line Element) data. The improvement on orbit prediction achieved 20-30% over the future 30 days period. The limitation of this hybrid approach is the requirement of at least 3 years of historical TLE data for the machine learning models, but could be overcome by creating synthetic orbital data from a similar space object. This hybrid model can be easily packaged into a software tool for space mission operation planning and facilitate mission autonomy.