Decarbonization dispatch for AI computing center parks considering task-level carbon traceability

The decarbonization dispatch of AI computing centers should not stop at controlling total park-level carbon emissions; it should also trace the energy sources, carbon-emission responsibility, and low-carbon labels of individual AI tasks. To address the limitation that existing studies mainly adopt park-average accounting and therefore cannot support task-level accounting, this paper proposes a decarbonization dispatch optimization method for AI computing center parks considering task-level carbon traceability. Taking the task execution trajectory as the main line, the proposed method links resource-pool startup and shutdown, IT power consumption, thermally inertial cooling, network storage, and energy-storage operation, thereby forming a closed relationship among the task ledger, energy-consumption ledger, and low-carbon ledger. Meanwhile, a unique low-carbon electricity claim and a label-violation CVaR constraint are introduced to control the risk of label failure under uncertain scenarios. Rolling-dispatch results on the Chenggong, Yunnan, China and the modified IEEE 118-bus system show that the proposed method can significantly improve task-level low-carbon performance with only a small increase in low-carbon attribute procurement and energy-storage throughput. In the Chenggong park case, the operating cost is reduced by 6.6%, and market-based park carbon emissions are reduced by 11.3%. The method provides a dispatch modeling basis for shifting low-carbon computing services from park-average accounting to task-level verification.