Data Center Model for Transient Stability Analysis of Power Systems

Overview

This arXiv preprint by Alberto Jimenez-Ruiz and Federico Milano (University College Dublin, 2025, arXiv:2505.16575) proposes a dynamic data-center load model for transient stability studies. The model has three main parts: a UPS between servers and grid that disconnects on voltage or frequency excursions and later reconnects, cooling represented as an induction motor (chiller compressors and pumps), and a pulsing load representing AI training transients.

The model is applied to the all-island Irish transmission system with data on connected data centers. As context, a fault in the Irish system caused a 204 MW drop in data-center demand, giving a RoCoF of 0.12 Hz/s and a frequency peak of 50.22 Hz. Irish data-center penetration is about 800 MW over a 7 GW system peak and is expected to reach 1.6 GW, or 30% of demand, by 2030. Cooling accounts for about 30% of total power of standalone data centers. Typical UPS switching margins are 0.2-0.3 Hz in frequency and 10% of nominal voltage.

In the case study, 300 MW of servers cooled by a 60 MW chiller, with UPS tripping at +/-0.1 pu or +/-0.3 Hz. With a 10 s disconnection time, the sudden reconnection pushed frequency below 49.7 Hz and re-disconnected the site, a cycle that can repeat. A data center made of UPSs with different reconnection times allows seamless reconnection with only moderate transients. Ramp-rate settings of compute load can soften AI-driven swings. Results are limited to the Irish system and the assumed model parameters.

Key points

1. A fault in the Irish system caused a 204 MW drop in data-center demand, giving a RoCoF of 0.12 Hz/s and a frequency peak of 50.22 Hz.
2. Irish data-center penetration is about 800 MW over a 7 GW system peak and is expected to reach 1.6 GW, or 30% of demand, by 2030.
3. Cooling accounts for about 30% of total power of standalone data centers and is modelled as an induction motor.
4. Typical UPS switching margins are 0.2-0.3 Hz in frequency and 10% of nominal voltage; outside either the UPS disconnects the loads.
5. Case: 300 MW of servers cooled by a 60 MW chiller, UPS tripping at +/-0.1 pu or +/-0.3 Hz; with a 10 s disconnection time, sudden reconnection pushed frequency below 49.7 Hz and re-disconnected the site, a cycle that can repeat.
6. A data center made of UPSs with different reconnection times allows seamless reconnection with only moderate transients.
7. The model has three parts: UPS, cooling as an induction motor, and a pulsing AI load.
8. Ramp-rate settings of compute load can soften AI-driven swings.

Using it in design

• Expose UPS transfer thresholds (voltage %, frequency Hz), return-to-mains delay and staggering across UPS modules as design parameters, and warn when all modules share identical thresholds and delays, since synchronised reconnection can re-trip the grid.
• Model chiller and pump motors as dynamic load in grid-study exports, and give the AI load a ramp-rate setting.
• For sites in small or islanded grids (data-center share of demand above about 10-30%), raise the priority of ride-through and staged-reconnection checks.

Who should read it

Electrical engineers studying the grid impact of large data centre loads, grid connection study specialists and those designing UPS parameters.

Publication details

Authors

Alberto Jimenez-Ruiz; Federico Milano

Publisher

arXiv preprint (University College Dublin)

Year

2025

Document number

arXiv:2505.16575

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2505.16575

Licence

CC BY-NC-ND 4.0 — share unaltered with attribution for noncommercial purposes

Citation