The explosive proliferation of distributed renewable energy exposes micro-clusters to millisecond-scale joint energy-and-information scheduling under rapid topology transients. We propose EB-STGA, an edge-batched spatio-temporal graph attention model that completes such scheduling within a 100 ms rolling horizon. By learning link delays and packet losses as edge weights and recalculating attention only over the perturbed edge set Δℰ_t, the method reduces computational complexity from O(|ℰ|) to O(|Δℰ|) while preserving privacy under the China 5G-Ready federated framework. Experiments on an extended IEEE-123 micro-cluster (120 V2G EVs, 60 battery units) show that EB-STGA cuts operational cost by 14.8%, limits voltage violations to 0.5%, and infers in 12 ms; zero-shot transfer to IEEE-33 yields <3% cost deviation. The results offer a new paradigm for real-time coordination of highly penetrated distributed resources in carbon-neutral power systems.