通信与信息技术

大规模的风、光新能源进入智算中心的电力保障后，“发电的间歇性”与“实时的算力供给”成为智算中心发电供需矛盾。现有调度方案存在维度割裂、算电耦合不足、碳管控停留在事后统计等问题。本文创新提出算力-电力-碳排放三方协同调度的总体框架，构建风光、算力、制冷注意力时序预测模型，提出“日前-日内-实时”三级递阶调度，并结合四级SLA业务分级实现弹性算力的精细化调节，形成全过程闭环优化体系。借助某平台搭建园区10 MW标准化仿真环境，与多种现有及前沿调度算法开展对比验证。结果表明：园区绿色电力就地消纳率由48.2%提高到85.7%，机房的PUE降低至1.18，单位算力碳排放降低40.2%，核心业务供电可靠性保持在99.999%。研究成果完善了源网荷储算冷一体化调度理论，可为“东数西算”算力枢纽低碳运行提供工程参考。

After large-scale integration of volatile wind and solar power into intelligent computing centers, the inherent intermittency of renewable generation has become a core supply–demand conflict against the real-time computing power demand.Existing dispatching approaches suffer from dimensional fragmentation, insufficient coupling between computing and power systems,and carbon management limited to post-hoc accounting. To address these issues, this paper proposes a novel three-party collaborative dispatching framework integrating computing power, electric power, and carbon emissions. A coupled attention-based time-series prediction model for renewable generation, computing load, and cooling demand is developed. A three-stage day-ahead/intra-day/real-time hierarchical dispatching strategy is designed, combined with four-level SLA-based elastic computing power refinement regulation, forming a full-process closed-loop optimization system. A standardized 10 MW simulation environment is built on the MATLAB/Simulink platform, and comparative validation is conducted against multiple existing and state-of-the-art dispatching algorithms. The results demonstrate that the on-site green power consumption rate increases from 48.2%to 85.7%, PUE decreases to 1.18, carbon emissions per unit of computing power are reduced by 40.2%, and the power supply reliability of core businesses remains at 99.999%. The research findings refine the integrated dispatching theory covering source, grid,load, storage, computing, and cooling, and provide an engineering reference for the low-carbon operation of “East–West Computing”hubs.