Data-Driven Real-Time Satellite Temperature Prediction in Thermal Vacuum Test Using Physics-Guided Long Short Term Memory

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Satellites are required to operate reliably under extreme space environments characterized by severe temperature variations and ultra-high vacuum conditions. Thermal vacuum tests (TVTs) are essential ground-based verification processes conducted prior to launch, during which large volumes of time-series data are collected over several weeks. As satellite systems become more complex, effectively utilizing these high-volume experimental data for real-time thermal state assessment remains a significant challenge.This study proposes a data-driven real-time satellite temperature prediction framework using a Long Short-Term Memory (LSTM) network trained on thermal vacuum test data. To improve prediction accuracy and physical consistency, a physics-guided data preprocessing strategy is introduced based on thermal engineering principles. By considering conductive and radiative heat transfer mechanisms, irrelevant variables are screened out, and physically meaningful temperature sensor data are selected to constrain the learning process.The proposed methodology is validated using actual thermal vacuum test data from a medium-sized satellite tested at the Korea Aerospace Research Institute (KARI). Prediction performance is evaluated through residual analysis by comparing measured temperatures with LSTM predictions. The results demonstrate that prediction accuracy improves as the training data volume increases: with more than 800,000 seconds of training data, prediction errors are reduced to within ±3°C, and further decrease to below ±1°C with extended training.The proposed approach enables real-time temperature estimation, early anomaly detection, and data recovery during satellite testing, demonstrating the practical applicability of machine learning to satellite hardware verification.

This publication has 4 references indexed in Scilit: