Mitigating Grid Stability Risks From Large AI Data Centres: A Three‐Mode Supervisory Control Framework for Centralized UPS Systems

ABSTRACT Hyperscale AI data centres concentrate hundreds of megawatts of inverter‐interfaced, voltage‐sensitive electronic load at single grid interconnection points, often in regions with limited short‐circuit capacity. During transmission faults, conventional uninterruptible power supply (UPS) systems permit near‐instantaneous load withdrawal that can produce aggregate power swings on the order of gigawatts, while broadband workload pulsing from AI training and inference pipelines can excite forced oscillations. This paper proposes a three‐mode supervisory control framework—a system‐level coordination of established control techniques—for centralized medium‐voltage UPS systems to mitigate these adverse grid impacts. Mode 1 shapes the facility's grid draw to attenuate broadband power pulsations, with the UPS battery energy storage system buffering the residual. Mode 2 enforces current‐limited reactive‐priority allocation during voltage sags—extended to asymmetric faults via positive‐sequence extraction—with a rate‐limited post‐fault recovery and a minimum‐draw policy that limits aggregate load withdrawal. Mode 3 optionally provides supervisory droop‐based frequency support, bounded by external frequency‐measurement and communication latency. A four‐controller simulation benchmark at an ultra‐weak node (short‐circuit ratio = 1.5) demonstrates zero unserved IT energy, 44% lower peak inverter current and 19% higher sustained point‐of‐common‐coupling voltage versus grid‐following baselines. Robustness sweeps across nine parameter dimensions confirm stable operation under varied grid strengths, fault types (balanced, SLG, LL, LLG), fault severities and workload profiles.