Integrated Low-Carbon Scheduling of Workload Shifting and Demand Response in AI Data Centers

AI data centers exhibit highly elastic electricity demand, where workload orchestration and grid demand response (DR) are tightly coupled and can compete in temporal allocation, resource utilization and carbon emission reduction. This paper proposes a coordinated computing–power low-carbon scheduling method that jointly optimizes workload shifting and DR participation for AI data centers. Firstly, a workload-shifting model with a closed-loop throttling–compensation structure is established to capture throttled execution, progress lag, and subsequent compensation over the task life cycle.

Secondly, a data-driven affine upper-envelope approach is developed to construct a conservative mapping from workload to nodal power, thereby preventing infeasible low-carbon schedules due to underestimated power and preserving a linear model structure. Then, a grid-oriented DR mechanism is incorporated with a conservative conversion scheme that translates throttling decisions into verifiable grid-side power reductions. To account for forecasting uncertainty in the workload-to-power mapping, a Wasserstein-distance-based distributionally robust optimization (DRO) model is formulated to jointly manage electricity cost and DR non-compliance risk.

Case studies show that the proposed method enables peak shaving and valley filling while maintaining service-level agreement (SLA) compliance, and they quantify how multidimensional workload constraints shape the practically deliverable DR capability, effectively supporting the decarbonization of AI data centers.