Data-driven in-orbit current and voltage prediction for LEO satellite lithium-ion battery state estimation

Abstract

Accurate estimation of the state of charge (SOC) of a satellite battery system is essential for determining missions and fault management design. However, in the case of low Earth orbit (LEO) satellites, it is difficult to continuously monitor the battery SOC from the ground due to non-contact duration in orbit. Therefore, to achieve accurate SOC estimation for the entire orbit, it is necessary to predict the battery power consumption. However, existing studies mostly use SOC estimation techniques that rely on real-time battery-related information, probability-based, and power budget-based techniques. Among these methods, the real-time onboard-based approach is not suitable for mission design and expansion because the status information is not available to the ground during the non-contact duration. Probability-based power state prediction and worst-case-based analysis are also not reliable during the non-contact duration. In this study, we propose current and voltage prediction by operation conditions utilizing bidirectional long short-term memory (Bi-LSTM) modeling, and, using the predicted data as input, the battery SOC is estimated using the unscented Kalman filter (UKF) algorithm considering battery degradation during the non-contact duration. The proposed technique is tested with in-orbit data of the KOMPSAT-3A satellite. As a result, the root mean squared error (RMSE) of current and voltage predictions achieved using the proposed technique are within about 1.7A and 0.2V, respectively. Additionally, the RMSE between the SOC estimated using the predicted current and voltage and the SOC calculated using the actual orbit data is found within 2%.