首页|期刊导航|综合智慧能源|考虑算力弹性伸缩的智算中心与电网协同优化运行关键技术综述

考虑算力弹性伸缩的智算中心与电网协同优化运行关键技术综述OA

Review of key technologies for coordinated optimal operation of intelligent computing centers and power grids considering elastic scaling of computing power

中文摘要英文摘要

针对智算中心作为新型电力系统中高功率密度、高波动关键负荷所引发的协同运行问题,沿着"任务—资源—能耗—电网"跨层级耦合链条,系统梳理了考虑算力弹性伸缩的智算中心与电网协同优化运行关键技术.将算力任务分为离线训练、在线推理和科学计算3类,分析了离线训练时间可平移与在线推理空间可迁移的差异化弹性特征,综述了物理能耗建模、时空负荷预测、异构算力池化调度、动态电压频率调整(DVFS)、断点续训和储能协同控制等关键技术,并总结了智算中心作为广义虚拟电厂参与需求响应与电力市场的联动机制;进一步基于深度强化学习将算电协同策略建模为马尔可夫决策过程,设计了涵盖"算随源动"时间维度策略与"算随电走"空间维度策略的云-边-端协同联动机制.研究表明:训练任务可通过任务重排、功率调节和断点续训等方式释放时间弹性,是削峰填谷的主力调节资源;推理任务可通过请求路由与跨区域迁移释放空间弹性,但受服务等级协议(SLA)约束呈低弹性特征;科学计算任务允许在特定时间窗口内错峰执行;设备调频与储能可补充快速功率响应,但整体调节能力受通信带宽、热安全、备用容量及设备折损等多重因素约束.Google碳感知算力调度、Microsoft电网友好型储能、冀北张家口电算协同虚拟电厂及宁夏中卫"源网荷储算"一体化等国内外工程实践已初步验证算力负荷参与新能源消纳、削峰填谷和辅助服务的技术可行性,但多数仍处于试点或示范阶段,规模化应用仍面临电算数据接口不统一、弹性边界难以可信量化、跨域安全约束耦合不足、多市场价值传导机制不完善及缺乏统一灵活性交易框架等瓶颈.未来应建立覆盖任务、设备、智算中心与电网的多时间尺度统一模型,在调度算法中引入硬件折损模型与算力中断代价以平衡电网需求响应红利与算力资产物理损伤,完善算力弹性评估、基线核验和电-算-碳协同机制,为"源网荷储算"融合运行提供支撑.

Intelligent computing centers,serving as high-power-density and highly fluctuating critical loads in new power systems,pose challenges to coordinated operation.To address these issues,key technologies for coordinated optimal operation of intelligent computing centers and power grids considering elastic scaling of computing power are systematically reviewed along the cross-layer coupling chain of"task—resource—energy consumption—power grid".Computing tasks are classified into three types:offline training,online inference,and scientific computing.The differentiated elasticity characteristics of temporal shiftability for offline training and spatial migratability in online inference are analyzed.Key technologies,including physical energy consumption modeling,spatiotemporal load prediction,heterogeneous computing resource pooling and scheduling,dynamic voltage and frequency scaling(DVFS),checkpoint-based training resumption,and coordinated energy storage control,are reviewed.In addition,the linkage mechanisms through which intelligent computing centers participate in demand response and electricity markets as generalized virtual power plants are summarized.Furthermore,computing-electricity coordination strategies are modeled as a Markov decision process based on deep reinforcement learning,and a cloud-edge-end coordinated linkage mechanism incorporating the temporal strategy of"computing follows energy sources"and the spatial strategy of"computing follows electricity"is designed.The results indicate that training tasks can release temporal elasticity through task reordering,power adjustment,and checkpoint-based training resumption,serving as the principal regulation resource for peak shaving and valley filling.Inference tasks can release spatial elasticity through request routing and cross-regional migration,but exhibit low elasticity due to service-level agreement(SLA)constraints.Scientific computing tasks can be executed during off-peak periods within specified time windows.Equipment frequency regulation and energy storage can provide supplementary rapid power response,although the overall regulation capability is constrained by multiple factors,such as communication bandwidth,thermal safety,reserve capacity,and equipment degradation.Engineering practices in China and abroad,including Google's carbon-aware computing scheduling,Microsoft's grid-friendly energy storage,the electricity-computing coordination virtual power plant in Zhangjiakou,northern Hebei,and the integrated"source-grid-load-storage-computing"project in Zhongwei,Ningxia,have preliminarily verified the technical feasibility of computing load participating in renewable energy consumption,peak shaving and valley filling,and ancillary services.However,most of these practices remain at the pilot or demonstration stage,and large-scale application still faces bottlenecks such as inconsistent electricity-computing data interfaces,difficulty in credibly quantifying elasticity boundaries,insufficient coupling of cross-domain security constraints,incomplete value transmission mechanisms across multiple markets,and the lack of a unified flexibility trading framework.Future efforts should establish a unified multi-timescale model covering tasks,equipment,intelligent computing centers,and power grids.Hardware degradation models and computing interruption costs should be incorporated into scheduling algorithms to balance the benefits of power grid demand response against physical damage to computing assets,and computing elasticity assessment,baseline verification,and electricity-computing-carbon coordinated mechanisms should be improved,thereby supporting integrated"source-grid-load-storage-computing"operation.

李彬;翟振杨;祁泽源;李淑静;陈宋宋;徐靖

华北电力大学 电气与电子工程学院,北京 102206华北电力大学 电气与电子工程学院,北京 102206华北电力大学 电气与电子工程学院,北京 102206华北电力大学 电气与电子工程学院,北京 102206中国电力科学研究院有限公司,北京 100192中国电力科学研究院有限公司,北京 100192

能源科技

智算中心算力弹性伸缩协同优化源网荷储算需求响应虚拟电厂算电协同

intelligent computing centerelastic scaling of computing powercoordinated optimizationsource-grid-load-storage-computingdemand responsevirtual power plantcomputing-electricity coordination

《综合智慧能源》 2026 (8)

44-56,13

国家电网有限公司科技项目(52060026007C-256-ZN)智能电网国家科技重大专项(2026ZD0811203)Science and Technology Project of State Grid Corporation of China(52060026007C-256-ZN)National Science and Technology Major Project on Smart Grid(2026ZD0811203)

10.3969/j.issn.2097-0706.2026.08.004

评论