A hybrid recurrent neural network and long short-term memory for simplified general perturbations-4 model in orbit propagation

• Main Author: Salleh, Nor ’Asnilawati
• Format: Thesis
• Language: English
• Published: 2022
• Subjects: T Technology (General); TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://eprints.utm.my/id/eprint/102399/1/NorAsnilawatiSallehPRAZAK2022.pdf; http://eprints.utm.my/id/eprint/102399/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:151646
• Institution: Universiti Teknologi Malaysia

Orbit propagation is one of the critical science tasks used to determine and forecast the position and velocity of orbiting space objects such as satellites, mission-related debris, rocket bodies, and others. Developing an accurate orbit propagation model is vital to ensure uninterrupted operational planning and prevent any disrupted collisions or disasters. However, using the current orbit propagation model has limitations, and these reduce the ability for long-term forecasting. It has errors depending on various aspects like measurement error, space environment information that constantly changes, inherent uncertainty in the data used, and errors in the data processing. Although classical time series methods such as Holt-Winters can improve the orbit propagator's accuracy and efficiency, it requires changes in the components' probability distribution, causing complexity and computational burden for end-user. However, this method can achieve maximum performance through integration with other approaches. Deep learning techniques, the new field of research within machine learning, are recently explored to analyse and improve the Simplified General Perturbations-4 (SGP4) Model, the orbit propagation model commonly used by space operators. The improved model should minimize errors and maintain accuracy even if the propagation span increases. Therefore, this study examined the Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) technique, a deep learning approach dealing with long-term time-series data. It can learn tasks and deal with complicated problems. Additionally, these learning techniques are a time series forecasting method that can improve models by capturing periodic data patterns by memorizing and learning from historical data. Thus, a hybrid RNN-LSTM SGP4 Model was developed. The performance and effectiveness of the improved model were evaluated and validated. As a result, this hybrid RNN-LSTM SGP4 Model improved more than 27% better than the SGP4 Model alone. It was also capable of being a reliable long-term time series forecasting model for space object data.