TL;DR
This paper introduces CA-PatchTST, a novel deep learning model for predicting solar array temperatures on satellites. By combining a unique decomposition technique with a Transformer-based architecture enhanced by cross-attention, CA-PatchTST achieves superior accuracy and robustness in forecasting long-term temperature trends, crucial for satellite health and autonomous operations.
What Problem Does This Paper Address?
Satellites operate in harsh environments, and their solar arrays are critical for power generation. Accurate temperature forecasting of these arrays is paramount for predictive maintenance, preventing failures, and ensuring mission reliability. However, satellite telemetry data for solar array temperatures presents significant challenges: it's high-dimensional, exhibits complex non-linear and non-stationary behavior due to orbital cycles and varying operational modes, and involves intricate inter-variable couplings. Traditional forecasting methods struggle with these characteristics, especially for long-term predictions, leading to error accumulation and reduced accuracy. This paper aims to bridge this gap by developing a method that can effectively capture these complex dynamics for reliable, multi-step temperature forecasting.
Methods
The core of the proposed solution is the CA-PatchTST model, which integrates several key innovations:
- Moving Average Decomposition: The raw temperature time-series is first decomposed into two components: a trend component (capturing slow-moving, long-term patterns like orbital cycles) and a residual component (representing short-term fluctuations and noise). This disentangles complex patterns and stabilizes the learning process.
- PatchTST Backbone: This architecture, adapted from the standard Transformer, processes the decomposed trend and residual components. It uses a "patching" strategy, dividing the time-series into smaller segments (patches) before applying self-attention. This significantly reduces computational complexity compared to standard Transformers, especially for long sequences, and improves the extraction of local temporal features. Crucially, it employs a channel-independent approach, meaning each sensor's data is processed separately within the encoder, preventing interference.
- Cross-Attention Mechanism: This is the key innovation for handling multivariate data. For a target sensor group (e.g., sensors on the solar wing body), its data acts as the "Query." Data from other sensor groups (e.g., BMSP structure, interface bracket) act as the "Key" and "Value." This allows the model to learn how the temperature of one component is influenced by the temperatures of others, capturing crucial inter-variable correlations and enabling information fusion.
- Output Fusion: The predictions from the trend and residual branches are combined through element-wise summation to produce the final multi-step forecast.
Datasets
The study utilizes the GOCE satellite solar array temperature telemetry dataset. This dataset comprises 16 temperature parameters collected from different locations on the GOCE satellite's solar array, spanning from March 2009 to June 2012. The data is sampled every 10 minutes and exhibits strong orbital periodicity, seasonal variations, and non-stationary behavior, making it a challenging benchmark for forecasting.
Key Insights and Ideas
- Decomposition is Key for Stability: Separating the temperature data into trend and residual components is crucial. The trend captures the predictable orbital cycles, while the residual captures the more erratic, short-term behaviors. This separation significantly improves the model's ability to handle non-stationarity and noise.
- Patching for Efficiency and Locality: The PatchTST's patching strategy is a clever way to make Transformer-like models more efficient for long time-series. It allows the model to focus on local patterns within patches while still capturing long-range dependencies by processing sequences of patches.
- Cross-Attention for Inter-Sensor Dynamics: The cross-attention mechanism is vital for understanding how different parts of the solar array influence each other's temperatures. The study's attention analysis shows that the model learns physically meaningful correlations, aligning with known thermal couplings between different components. This is a significant step beyond models that treat sensors independently.
- Synergy of Components: The ablation studies clearly demonstrate that neither the decomposition nor the cross-attention mechanism alone is sufficient. Their combination in CA-PatchTST leads to superior performance, highlighting their complementary roles in capturing both long-term trends and short-term, inter-dependent dynamics.
Key Results
- Superior Accuracy: CA-PatchTST consistently outperformed several state-of-the-art baselines (including DLinear, SegRNN, Informer, TimesNet, and iTransformer) across multiple forecast horizons (24h, 72h, and 120h).
- Significant Error Reduction: The model achieved notably low MAPE values: 9.26% at 24h, 13.68% at 72h, and 24.67% at 120h, significantly surpassing alternatives. For example, at 120h, its MAPE was 24.67% compared to 47.22% for SegRNN.
- Robustness and Stability: The model demonstrated high stability, with error curves remaining close to zero, indicating reliable performance even over extended forecast periods.
- Computational Efficiency: CA-PatchTST achieves a favorable balance between performance and efficiency, boasting a compact architecture (4.9M parameters), fast inference (10.5ms/batch), and low memory consumption (2.4GB), making it suitable for resource-constrained satellite systems.
Limitations and Future Work
- Future Work: The authors plan to focus on model lightweighting (pruning, quantization, knowledge distillation) and online adaptation (continual learning, anomaly-aware updates) to optimize the model for operational deployment and maintain performance under evolving environmental conditions.
- Limitations: While the paper highlights the model's effectiveness, the inherent complexity of satellite telemetry data means that perfect prediction is an ongoing challenge. The paper focuses on temperature, and future work could extend to other critical telemetry parameters. The computational efficiency, while good, could still be a limiting factor for extremely resource-constrained platforms, motivating the lightweighting efforts.
Summary updated