Analysis on AI Training Data Center Siting and Demand Response in Power System Operation

AI services are driving rapid growth in electricity demand from AI training data centers (AI DCs). With increasing renewable penetration, AI DC siting and flexibility can change congestion, reserves, and curtailment. We develop a MILP-based unit commitment (UC) model that embeds a two-state (idle/training) AI DC load with a fixed training-energy requirement and minimum dwell-time constraints. The AI DC schedule is co-optimized with thermal UC, reserves, and DC power-flow limits. Using a reduced Korean system model, we evaluate four scenarios combining siting (Metropolitan vs. Jeonnam) and operation (constant vs. demand-responsive) for 2030 and 2035 using yearly rolling-horizon simulations. In 2035, the Jeonnam siting plus demand-response case reduces curtailment by 30.92% and operating cost by 4.75% (KRW 56.7 billion) relative to the baseline. The framework supports quantitative grid-impact assessment and policy design for large AI DC loads.