Retrieval-augmented patch generation for geosynchronous satellite status forecasting.

Accurate geosynchronous satellites status forecasting is essential for improving space situational awareness and supporting downstream tasks such as maneuver detection and intent inference. However, current models face difficulty modeling high-orbit observation data due to its strong non-stationary and weak periodic structure, which hinders effective long-range dependency learning and adaptation to evolving orbital dynamics. To address these challenges, we propose RAPG, a Retrieval-Augmented Patch Generation framework for geosynchronous satellite status analysis, including status forecasting and intent classification tasks. RAPG integrates two synergistic components: (1) an adaptive frequency-domain patching mechanism that dynamically segments satellites' status time series according to their dominant spectral characteristics, enabling the model to capture both short-term and long-term temporal structures; and (2) a retrieval-augmented generation module that accesses historically similar patches from the training set, providing explicit external memory for reconstructing weakly periodic and non-repetitive motion patterns. By linking adaptive patching with retrieval-based augmentation, RAPG enhances temporal representation learning and generalization across diverse orbital conditions. Experiments on two synthetic datasets and one real-world dataset show that RAPG consistently surpasses strong baselines in both status forecasting and intent classification. RAPG yields the lowest forecasting errors and achieves an F1-score of 0.9461 and an accuracy of 0.9443 for intent classification, outperforming baseline methods. These results confirm that RAPG provides a robust and scalable framework for real-time analysis of geosynchronous satellites, offering significant potential for enhancing remote sensing-based space domain awareness and dynamic orbital intelligence.