Low‑Earth‑orbit satellites move quickly, forcing frequent handovers of user connections. Traditional handover settings—fixed thresholds for signal strength and timing—cannot keep up with changing signals and satellite paths. A new framework uses reinforcement learning on the policy of a satellite network to choose the right moment and target for each handover. The policy (PPO‑SHD) takes into account the signal power received, which satellites are connected to ground stations, how long each satellite stays above a minimum elevation angle, and the geometry of the links.

In simulations, the PPO‑SHD policy keeps the network available 99.32 % of the time when the minimum elevation angle is set to 40°. That is 1.67 percentage points higher than the best tuned fixed‑threshold method (O‑HT) and 1.41 p.p. higher than a method that simply picks the satellite with the shortest delay (LD). In the same trials, the policy cuts the cumulative radio‑link‑failure time by 85.7 % compared with O‑HT and by 85.0 % compared with LD.