Data Center Model for Transient Stability Analysis of Power Systems

TL;DR

The paper addresses the stability risks introduced by data centers in power systems by proposing a dynamic DC load model that captures three key features: (1) a UPS between IT loads and the grid, (2) cooling load modeled as an induction motor, and (3) pulsing AI workloads. The model integrates detailed server-side dynamics (CPU/GPU loads with bursts and smoothing), a realistic cooling and energy-storage system, and robust UPS reconnection logic, and it is validated on the all-island Irish transmission system using real DC data. Case studies show that DCs can trigger strong transients and, if reconnection is not carefully managed, lead to grid instability or flapping; strategies such as gradual reconnection or load segmentation can mitigate these effects. The results highlight the importance of incorporating DC-specific dynamic behavior into transmission models and suggest directions for control strategies to preserve grid stability while enabling DC growth. $p_{DC}$ is decomposed into $p_{Cooling}+p_{ZIP}+p_{IT}$, and the UPS dynamics involve $v_i$, $ heta_i$, and mode-dependent power flows, all of which are essential for accurate transient simulations.$

Abstract

The rising demand of computing power leads to the installation of a large number of Data Centers (DCs). Their Fault-Ride-Through (FRT) behavior and their unique power characteristics, especially for DCs catered to Artificial Intelligence (AI) workloads, pose a threat to the stability of power systems. To ensure its stability, it is required accurate models of the loads involved. Here we propose a dynamic load model that properly captures the behaviour of DCs. Its three most defining features are the use of an Uninterrupted Power Supply (UPS) which sits between the server load and the grid, the cooling load represented by an induction motor, and a pulsing load that represents the transients caused by contemporary DCs with significant AI workloads. The features of the proposed model and its impact on the dynamic performance of transmission systems are illustrated through a model of the all-island Irish transmission system and real-world data of the DCs currently connected to this system.

Figures (15)

• Figure 1: Frequency and RoCoF for Kellystown - Woodland fault kerciImpactConverterbasedDemand2024.
• Figure 2: Batched load.
• Figure 3: Power consumption of transient CPU scaling driver: intel_pstate.
• Figure 4: load.
• Figure 5: internal structure.