首页|期刊导航|综合智慧能源|算电协同架构下基于MCP的多智能体协同机制综述

算电协同架构下基于MCP的多智能体协同机制综述OA

Review of MCP-based multi-agent synergistic mechanisms in computing-electricity coordination architecture

中文摘要英文摘要

随着大语言模型驱动的多智能体系统在电力领域规模化部署,算力设施负荷呈现显著的阵发性与潮汐性波动特征,与新型电力系统安全稳定运行之间的矛盾日益突出.2025年我国算力中心用电量已达196 TW·h,预计2030年将突破700 TW·h,算电协同市场规模将超万亿元;叠加异构智能体交互接口标准缺失,导致算力资源配置失衡、新能源消纳受阻、跨域调度存在通信孤岛等多重行业痛点.针对物理资源错配与数字交互割裂的双重核心矛盾,系统梳理了算电协同、多智能体交互与模型上下文协议(MCP)3个维度的国内外前沿研究进展.在物理机制层面,解构了算力、电力、碳排放多维耦合运行机制,归纳了电力安全校核与动态碳足迹量化两类技术体系,阐述了基于碳排放流理论的实时碳责任溯源方法及动态低碳标签模型,揭示了延迟容忍型计算任务的时空平移与跨域迁移潜力;在通信架构层面,剖析了MCP的3层有状态通信架构与标准化交互原语的技术优势,对比传统网状对接方案适配成本过高的短板;在协同机制层面,总结了物理约束驱动的双层跨域协同架构设计思路,梳理了算力路由、内生安全防护、多方收益均衡分配等配套技术进展,并结合南方电网电碳算协同运营系统等算力枢纽工程场景归纳了现有落地范式.综合研判指出,当前研究在弱网环境下智能体状态同步、大规模异构集群协同调度、跨信任域动态权限管控等方面仍存在技术瓶颈,未来应聚焦标准化评估体系构建、跨域协同底层协议研制、多目标联合优化算法突破等方向,可为新型电力系统下算电协同多智能体的标准化与安全化落地提供理论支撑,推动数字经济转型与能源低碳转型的深度融合与协同发展.

With the large-scale deployment of large language model-driven multi-agent systems in the power sector,computing facility loads exhibit pronounced bursty and tidal fluctuation characteristics,resulting in increasingly prominent conflicts with the secure and stable operation of new power systems.Electricity consumption by China's computing centers reached 196 TW·h in 2025 and is expected to exceed 700 TW·h by 2030,while the computing-electricity coordination market is projected to exceed 1 trillion yuan.Combined with the lack of standardized interaction interfaces for heterogeneous agents,these trends lead to multiple industry challenges,including imbalanced computing resource allocation,impeded renewable energy consumption,and communication silos in cross-domain scheduling.To address the dual-core conflicts of physical resource misallocation and fragmented digital interaction,the latest research progress in China and abroad is systematically reviewed across three dimensions:computing-electricity coordination,multi-agent interaction,and the Model Context Protocol(MCP).At the physical mechanism level,the multidimensional coupled operating mechanism among computing power,electricity,and carbon emissions is deconstructed.Two technical systems,namely power security verification and dynamic carbon footprint quantification,are summarized.The real-time carbon responsibility tracing method based on carbon emission flow theory and the dynamic low-carbon labeling model are described,and the potential for spatiotemporal shifting and cross-domain migration of delay-tolerant computing tasks is revealed.At the communication architecture level,the technical advantages of MCP's three-layer stateful communication architecture and standardized interaction primitives are analyzed,while highlighting the excessive adaptation costs associated with conventional mesh-based integration schemes.At the synergistic mechanism level,the design approach of a two-tier cross-domain synergistic architecture driven by physical constraints is summarized,together with advances in supporting technologies such as computing power routing,intrinsic security protection,and balanced allocation of benefits among multiple parties.Additionally,existing implementation paradigms are summarized in combination with computing hub engineering scenarios,such as the electricity-carbon-computing synergistic operation system of the China Southern Power Grid.A comprehensive assessment indicates that current research still faces technical bottlenecks in agent-state synchronization in weak-grid environments,synergistic scheduling of large-scale heterogeneous clusters,and dynamic access control across trust domains.Future efforts should focus on establishing standardized evaluation systems,developing underlying protocols for cross-domain synergy,and achieving breakthroughs in multi-objective joint optimization algorithms.These directions can provide theoretical support for the standardized and secure implementation of multi-agent computing-electricity coordination in new power systems and promote deeper integration and coordinated development of the digital economy transition and the low-carbon energy transition.

李彬;申楠;徐靖;陈宋宋

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

算电协同算力中心大语言模型多智能体系统模型上下文协议跨域调度负载迁移碳排放流

computing-electricity coordinationcomputing centerlarge language modelmulti-agent systemmodel context protocolcross-domain schedulingload migrationcarbon emission flow

《综合智慧能源》 2026 (8)

16-28,13

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.2097-0706.2026.08.002

评论