首页|期刊导航|综合智慧能源|智算中心柔性负荷建模与算电协同调控关键技术及展望

智算中心柔性负荷建模与算电协同调控关键技术及展望OA

Key technologies and prospects for flexible load modeling and computing-electricity coordinated regulation in intelligent computing centers

中文摘要英文摘要

针对智算中心柔性负荷来源分散、跨层耦合复杂及多园区电-算-碳协同缺乏统一框架等问题,以任务负荷、算力资源、设施设备、并网功率跨层传递链条为主线,系统综述了智算中心柔性负荷建模与算电协同调控研究.在方法上,首先区分离线、在线及混合3类工作负载的功率特征、可用调节动作与恢复反弹风险,归纳计算侧、电源侧和热管理侧柔性负荷的形成机制及其响应、持续与恢复的时间尺度差异,并明确多类柔性资源在协同调用中的先后次序.其次,总结基线负荷、柔性包络与不确定性预测方法,比较分时电价、可再生能源出力及爬坡指令等外部信号驱动下的典型需求响应模式,并从电网、设施和用户三侧归纳可交付容量、储能退化及服务质量代价等评价指标.随后,进一步分析多智算园区源-荷-储-算耦合架构、任务时空迁移机制、动态碳核算方法及电价、绿电与碳信号的协同传导过程,并对不同协同机制的驱动信号、优化目标与关键约束进行了横向比较.在结果与结论方面,智算中心柔性具有跨层级、多时间尺度和状态相关等特征,可靠参与电网调节的关键在于将内部可调潜力转化为可预测、可交付和可核验的聚合响应:可预测对应调用前的容量与置信度评估,可交付对应调用中的约束兑现能力,可核验对应调用后基于统一基线的量效核实,三者共同构成预测、承诺、调用、计量结算、滚动更新的工程闭环.同时,智算中心与传统数据中心在功率密度、同步通信及热安全边界等方面存在差异,部分传统结论仍需针对智算业务重新验证.最后,归纳了多能耦合、电网安全、规划运行协同、低碳价值计量、电网-算力网络跨区域协同及人工智能原生自主控制等未来重点发展方向,并指出现有研究在工程证据等级、负效应归纳及跨园区风险协同治理方面仍存在不足,为新型电力系统中算力负荷柔性挖掘与需求响应机制提供参考.

To address the dispersed sources of flexibility,complex cross-layer coupling,and the lack of a unified framework for electricity-computing-carbon coordination across multiple intelligent computing parks,a systematic review of research on flexible load modeling and computing-electricity coordinated regulation in intelligent computing centers is conducted,with the cross-layer transmission chain from workloads,computing resources,facility equipment,and grid-connected power as the mainline.The power characteristics,available regulation actions,and recovery-related rebound risks of offline,online,and hybrid workloads are distinguished.The formation mechanisms of computing-side,power-side,and thermal-management-side flexibility are summarized,and the differences in response,duration,and recovery timescales as well as the priority relationships among different flexibility resources are clarified.Then,baseline load modeling,flexibility envelope characterization,and uncertainty-aware prediction methods are reviewed.Typical demand response modes driven by external signals such as time-of-use electricity prices,renewable energy output,and ramping commands are compared,and evaluation indicators including deliverable capacity,energy storage degradation,and quality-of-service costs are summarized from the perspectives of power grids,facilities,and users.Subsequently,the source-load-storage-computing coupling architecture,spatiotemporal workload migration mechanisms,dynamic carbon accounting methods,and coordinated transmission processes of electricity price,renewable energy,and carbon signals in multiple intelligent computing parks are analyzed.Different coordination mechanisms are further compared in terms of driving signals,optimization objectives,and key constraints.The review indicates that the flexibility of intelligent computing centers exhibits cross-layer,multi-timescale,and state-dependent characteristics.The key to reliable participation in grid regulation lies in transforming internal adjustable potential into predictable,deliverable,and verifiable aggregated responses.Specifically,predictability corresponds to capacity and confidence evaluation before dispatch,deliverability refers to the capability of satisfying operational constraints during dispatch,and verifiability represents the performance validation based on unified baselines after dispatch.Together,these three elements establish an engineering closed loop covering prediction,commitment,dispatch,measurement and settlement,and rolling updates.Meanwhile,intelligent computing centers differ from conventional data centers in terms of power density,synchronized communication requirements,and thermal safety boundaries,indicating that some existing conclusions require further validation for intelligent computing workloads.Future key development directions are summarized,including multi-energy coupling,power grid security,coordinated planning and operation,low-carbon value measurement,cross-regional coordination between power grids and computing networks,and AI-native autonomous control.Existing challenges are also identified in engineering-level validation,explicit consideration of adverse effects,and cross-park risk coordination,providing references for exploiting computing load flexibility and developing demand response mechanisms in new power systems.

朱继忠;杨晓宁;张迪;王晞罗

华南理工大学 电力学院,广州 510641华南理工大学 电力学院,广州 510641太原理工大学 电气与动力工程学院,太原 030024华南理工大学 电力学院,广州 510641

能源科技

智算中心柔性负荷需求响应算电协同低碳电力调度

intelligent computing centerflexible loaddemand responsecomputing-electricity coordinationlow-carbon power dispatch

《综合智慧能源》 2026 (8)

1-15,15

国家自然科学基金项目(U25B6016)National Natural Science Foundation of China(U25B6016)

10.3969/j.issn.2097-0706.2026.08.001

评论