Dynamic Link Prediction of Space Conjunction

The conjunction prediction between Resident Space Objects (RSOs) is crucial for Space Situational Awareness (SSA) and Space Traffic Management (STM). With the increase of space objects, it poses a computational challenge to existing monitoring systems. In this work, we propose a novel end-to-end architecture that frames conjunction event as a dynamic network link prediction task. We generate a dataset to validate our approach using the public in-orbit data. Then, we model it as a discrete-time dynamic complex network, using nodes to represent RSOs and links for their conjunction events, and analyze the centrality, clusters, and evolution of this network. Finally, we compare various link prediction methods, including heuristics, static graph neural networks, and dynamic graph neural networks. The results demonstrate that the GC-LSTM model, which uses the position and velocity of RSOs as node features, achieves Mean Reciprocal Rank (MRR) and AUC (Area Under the receiver operating characteristic Curve) scores of 0.760 and 0.944, respectively. The proposed framework is 10 times faster than traditional numerical methods, can effectively and Quickly predict global and massive conjunction events.