Design of a Transformer-GRU-Based Satellite Power System Status Detection Algorithm

Guoqi Xie, Xinhao Yang, Jiayu Zhao, Zhou HuangBatteries 2025 · DOI

TL;DR

This paper introduces an enhanced Transformer-GRU model (TF-GRU) that integrates a Graph Attention Network (GAT) to better model the spatial relationships between battery cells in satellite power systems. This novel approach significantly improves the accuracy and efficiency of detecting anomalies and predicting the health status of these critical systems, addressing key limitations of existing methods.

What Problem Does This Paper Address?

Satellite power systems are crucial for mission success, yet they experience the highest failure rates among all spacecraft subsystems. A significant portion of these failures are attributed to energy storage battery packs. Existing methods for monitoring battery health often struggle with:

  • Inadequate spatial correlation modeling: Battery packs consist of interconnected cells, and faults can propagate between them. Standard Transformer models, while good at temporal dependencies, don't explicitly capture these physical, spatial relationships.
  • Computational inefficiency: Traditional Transformer decoders can be too slow for real-time satellite operations.
  • Difficulty with long-term temporal patterns: Satellite data often exhibits periodic variations (like eclipse cycles) and complex fault evolution that are challenging to model precisely.
  • Data scarcity and imbalance: Fault data is rare, making it difficult for models to learn subtle fault patterns.

This paper aims to overcome these challenges by developing a more robust and efficient algorithm for detecting the status of satellite power systems.

Methods

The core innovation is an improved Transformer-GRU model (TF-GRU) that combines the strengths of several architectures:

  • Transformer Encoder with GAT: A Graph Attention Network (GAT) is embedded within the Transformer encoder. This allows the model to explicitly learn and represent the spatial correlations and dependencies between individual battery cells, treating them as nodes in a graph. The GAT dynamically adjusts attention weights to capture how faults might propagate through the battery pack.
  • GRU Decoder: The conventional Transformer decoder is replaced with a Gated Recurrent Unit (GRU). GRUs are known for their efficiency and ability to capture short-term temporal dependencies, making them suitable for real-time processing while still maintaining temporal dynamics.
  • Spatiotemporal Feature Fusion: The architecture effectively fuses spatial features learned by the GAT-enhanced encoder with temporal features captured by the GRU decoder. This allows for a comprehensive understanding of the satellite power system's state.
  • Input Features: The model utilizes voltage, temperature, and temporal features as input.

Datasets

The study uses telemetry data collected from a geostationary orbit (GEO) communication satellite. This data includes voltage and temperature measurements from an onboard energy storage battery system.

  • Data Preprocessing: Abnormal data points were excluded, missing data was imputed, and the data was normalized.
  • Dataset Split: The data was chronologically partitioned into training (71,722 samples) and testing sets (4490 samples). The test set specifically captured complete charge/discharge cycles.
  • Fault Simulation: A simulated scenario was used to generate fault data, specifically a failure in the Battery Charge Regulator (BCR) leading to persistent CC charging and abnormal voltage/temperature increases.

Key Insights and Ideas

  • Explicitly modeling spatial relationships is crucial: The GAT's ability to capture inter-cell dependencies is a key differentiator, moving beyond treating the battery pack as a monolithic unit. This is vital for understanding fault propagation.
  • Hybrid architectures offer the best of both worlds: Combining the long-range dependency modeling of Transformers with the efficient temporal processing of GRUs, and adding spatial awareness with GATs, creates a powerful and practical solution.
  • Real-time performance is achievable: Replacing the Transformer decoder with a GRU significantly reduces computational complexity, making the model suitable for on-orbit applications.
  • Subtle anomalies can be detected early: The model's residual analysis allows for the detection of deviations from normal behavior, potentially flagging faults before they become critical.

Key Results

  • Superior Trend Prediction: The improved TF-GRU model achieved significantly lower Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) for both voltage and temperature prediction compared to the conventional Transformer and standard TF-GRU models.
  • Voltage: MAPE of 0.1712% and RMSE of 0.1729 V (vs. Transformer: 0.7776% MAPE, 0.7179 V RMSE; Standard TF-GRU: 0.6308% MAPE, 0.6108 V RMSE).
  • Temperature: MAPE of 1.1410% and RMSE of 0.2954 °C (vs. Transformer: 1.8406% MAPE, 0.4142 °C RMSE; Standard TF-GRU: 1.8780% MAPE, 0.4352 °C RMSE).
  • Accurate Fault Detection: The model successfully identified a simulated BCR fault with a detection delay of less than 2 seconds and a false alarm rate of 1.54%. It accurately flagged abnormal voltage increases that baseline models missed.
  • Statistical Validation: A Mann-Whitney U test confirmed the statistical significance of the improved voltage prediction performance.

Limitations and Future Work

  • Data Scarcity: The inherent scarcity of satellite telemetry data, especially fault data, remains a challenge. Future work could explore data augmentation and transfer learning.
  • Hardware Deployment: Deploying the model on resource-constrained satellite hardware (like FPGAs) might require further optimization through model compression or quantization.
  • Dynamic Connectivity: Future research should address dynamic variations in connection density due to thermal coupling and graph structural drift from battery aging.
  • Extreme Scenarios: The current study does not fully encompass extreme operating conditions (e.g., severe over/undercharging). Future work will validate performance in these cases using physics-based data augmentation.
  • Generalization: While the model shows promise for generalization across battery types, further investigation into sparse connectivity scenarios for GATs is needed.
  • Multimodal Data Synchronization: Challenges in synchronizing data from different sensors with varying frequencies will be addressed with advanced techniques like cross-modal temporal attention.

Summary updated

Abstract

The health state of satellite power systems plays a critical role in ensuring the normal operation of satellite platforms. This paper proposes an improved Transformer-GRU-based algorithm for satellite power status detection, which characterizes the operational condition of power systems by utilizing voltage and temperature data from battery packs. The proposed method enhances the original Transformer architecture through an integrated attention network mechanism that dynamically adjusts attention weights to strengthen feature spatial correlations. A gated recurrent unit (GRU) network with cyclic structures is innovatively adopted to replace the conventional Transformer decoder, enabling efficient computation while maintaining temporal dependencies. Experimental results on satellite power system status detection demonstrate that the modified Transformer-GRU model achieves superior detection performance compared to baseline approaches. This research provides an effective solution for enhancing the reliability of satellite power management systems and opens new research directions for future advancements in space power system monitoring technologies.
BibTeX
@inproceedings{xie2025design,
  author = {Guoqi Xie and Xinhao Yang and Jiayu Zhao and Zhou Huang},
  title = {{Design of a Transformer-GRU-Based Satellite Power System Status Detection Algorithm}},
  booktitle = {Batteries},
  year = {2025},
  doi = {10.3390/batteries11070256},
}
Figure