Inter-satellite links (ISLs) in large LEO constellations vary quickly with time,
and their rapid change makes future link-state prediction a practical concern.
Predicting ISLs with reasonable accuracy benefits routing decisions, resource
scheduling, and overall network stability. Existing methods are still limited in
several ways. Rule-based and physics-based methods build on hand-crafted
assumptions, which become unreliable once links flicker frequently or topologies
grow heterogeneous. Static graph neural networks describe the relations within a
single snapshot, but their formulation does not represent how ISLs evolve from
one moment to the next. Many recent dynamic graph models stay close to
interpolation, and they exploit physical and temporal satellite attributes only
in part, which weakens both node representations and prediction accuracy. To
address these gaps, we develop an extrapolation-oriented dynamic link prediction
framework with three contributions: (1) we shift the task from interpolation
toward future-state extrapolation, which makes proactive management possible;
(2) we extract physics- and temporal-aware node features by combining orbital
dynamics with historical ISL information; (3) we restrict prediction to a
candidate edge set, so the computation stays manageable while the method remains
scalable. In experiments, the proposed method outperforms existing baselines on
accuracy, efficiency, and scalability, suggesting that it can serve as a useful
tool for managing large-scale LEO networks in practice.