DC-CLM：面向AI数据中心动态的WECC复合负荷模型扩展
DC-CLM: Extending the WECC Composite Load Model for AI Data Center Dynamics
查看机构详情
- Louisiana Tech University（路易斯安那理工大学）
机构由 AI 辅助整理，请以论文原文为准。
浏览论文内容
中文总结 AI 辅助
本文提出DC-CLM，扩展WECC复合负荷模型以表征AI数据中心动态，涵盖UPS、冷却及工作负载状态，仿真表明训练/推理较空闲显著影响电压和功率变化。
中文摘要 AI 辅助
AI驱动数据中心的快速增长引入了传统复合负荷模型未明确表示的负荷行为。本文提出DC-CLM，一种面向工作负载的WECC复合负荷模型扩展，该模型纳入了UPS支撑的IT负荷、混合电机/变频器冷却负荷、辅助负荷以及训练、推理和空闲工作负载曲线。基于规则的监督状态机表示电网、电池和柴油运行状态，并捕获电压恢复后的临时IT负荷隔离和依赖工作负载的恢复过程。该模型在MATLAB/Simulink中使用正序相量域仿真实现，并在故障引起的延迟电压恢复（FIDVR）条件下进行评估。对于具有100 MW传统复合负荷和200 MW数据中心负荷的研究系统，长时间电压恢复主要由传统堵转电机热动态主导，而故障清除后前100 ms强烈依赖工作负载。训练和推理产生约50-52 mpu电压变化和751-755 MW有功功率变化，而空闲运行约为36 mpu和654 MW。结果表明，将数据中心特定负荷动态纳入输电级稳定性研究具有价值，同时强调未来需要基于测量的验证和更详细的变流器级建模。
英文摘要
The rapid growth of AI-driven data centers is introducing load behaviors that are not explicitly represented in conventional composite load models. This paper presents DC-CLM, a workload-aware extension of the WECC composite load model that incorporates UPS-supported IT demand, mixed motor/VFD cooling loads, auxiliary demand, and training, inference, and idle workload profiles. A rule-based supervisory state machine represents grid, battery, and diesel operating states and captures temporary IT-load isolation and workload-dependent restoration following voltage recovery. The model is implemented in MATLAB/Simulink using positive-sequence phasor-domain simulation and evaluated under fault-induced delayed voltage recovery (FIDVR) conditions. For a study system with a 100~MW conventional composite load and a 200~MW data-center load, the long-duration voltage recovery is governed mainly by conventional stalled-motor thermal dynamics, while the first 100~ms after fault clearance is strongly workload-dependent. Training and inference produce approximately 50--52~mpu voltage variation and 751--755~MW active-power variation, compared with approximately 36~mpu and 654~MW for idle operation. The results demonstrate the value of incorporating data-center-specific load dynamics into transmission-level stability studies while highlighting the need for future measurement-based validation and more detailed converter-level modeling.
