Forecast-to-Realized Carbon-Aware Dispatch Evaluation for AI Data Centers with Constrained Workload Migration and Battery Storage

AI data centers are emerging grid-interactive loads at the interface of information demand and power-network operation. This paper asks how much carbon-aware flexibility remains after a forecast-optimized schedule is settled on realized workload, weather, price, and carbon profiles. We formulate a forecast-to-realized dispatch-evaluation architecture for three geographically distributed AI data centers.

The carbon-aware dispatch layer schedules forecast-based workload routing, PV use, BESS operation, grid import, and aggregate peak import. The realized-value accounting layer replays routing as workload shares, recomputes demand on realized profiles, clips PV by realized availability, preserves storage actions, adds nonnegative balancing-grid import, and reports cost, carbon, peak, regret, service, and benchmark feeder metrics. The experiments use public Energy-Charts and Open-Meteo signal shapes with a reproducible scenario workload.

In the constrained main case, coordinated dispatch reduces cost, carbon emissions, and aggregate peak import by 3.33%, 4.00%, and 6.63%. Relative to a strict same-asset no-migration control, enabling migration adds 1.45% cost and 1.80% carbon reduction on the main day for the same BESS asset stack. Multi-day public-signal replays and paired forecast-error trials show that the reductions persist conditionally and that lower-error updated forecasts reduce regret relative to the perfect-information benchmark.

Sensitivity, service, trace-replay, tariff, carbon-signal, feeder-proxy, and scalability checks are interpreted as operating-regime checks rather than site-specific assessment.