Wide-Area Power System Oscillations from Large-Scale AI Workloads

Overview

The work develops a new dynamic power-profiling approach for AI-centric datacenter loads and analyses their impact on grid operations, particularly their potential to induce wide-area oscillations. It characterises the periodic stochastic power fluctuations inherent to large-scale AI workloads during training and fine-tuning, driven by GPU architecture. These sustained, large power fluctuations, unlike conventional load ramping, act as persistent forcing inputs capable of interacting with and amplifying local and inter-area oscillation modes.

Training model parameters: baseline fluctuation frequency U(0.5, 1.5) Hz, cycle-duration sigma 0.1, compute-phase share 55%-80% of each cycle, and an average 30% demand reduction in synchronisation phases (mu_Delta = 0.3). The fine-tuning model uses a dominant frequency U(0.3, 0.7) Hz and compute fraction 0.7-0.9. Baseline WECC case: 60.8 GW base load, system inertia 2.7, 7 datacenters of 1 GW each (11.5% penetration), fluctuation band 0.5-1.5 Hz, with 80% of each site's capacity modelled as steady load and 20% as fluctuating AI load.

Eigen-analysis of the WECC system found unstable modes at 0.852, 1.06 and 1.44 Hz; higher-penetration scenarios of 23% and 34.5% were also simulated. Across 20 stochastic runs the response consistently shows a dominant oscillation around 1 Hz, and narrower workload frequency ranges give stronger, more repeatable oscillations. Direct suppression of datacenter load fluctuation is substantially more effective than indirect system-level measures such as added damping.

The finding is that AI datacenters behave as persistent, frequency-selective forcing sources rather than ramping loads, and oscillation severity depends on fluctuation frequency band, single-site size, siting, system inertia and penetration. Limits: the stochastic model parameters are assumptions, not from a measured single-facility profile; the simulation is on two test systems. It is a preprint.

Key points

1. Non-computational datacenter components (cooling, networking, storage) are cited as typically 40-50% of total consumption, and AI workloads as roughly 10-20% of datacenter electricity today.
2. Training model parameters: baseline fluctuation frequency U(0.5, 1.5) Hz, cycle-duration sigma 0.1, compute-phase share 55%-80% of each cycle, and an average 30% demand reduction in synchronisation phases (mu_Delta = 0.3).
3. The fine-tuning model uses a dominant frequency U(0.3, 0.7) Hz and compute fraction 0.7-0.9.
4. Baseline WECC case: 60.8 GW base load, system inertia 2.7, 7 datacenters of 1 GW each (11.5% penetration), fluctuation band 0.5-1.5 Hz; 80% of each site as steady load and 20% as fluctuating AI load.
5. Eigen-analysis of the WECC system found unstable modes at 0.852, 1.06 and 1.44 Hz; higher-penetration scenarios of 23% and 34.5% were also simulated.
6. Across 20 stochastic runs the response consistently shows a dominant oscillation around 1 Hz; narrower workload frequency ranges give stronger, more repeatable oscillations.
7. Direct suppression of datacenter load fluctuation is substantially more effective than indirect system-level measures such as added damping.

Using it in design

• Model AI data-centre load as a persistent periodic forcing (0.3-1.5 Hz band, about 20% of peak fluctuating) rather than a ramp, and add oscillation analysis to the grid-interconnection study.
• Suppress fluctuation on site (UPS, battery, workload smoothing); direct suppression is far more effective than system-level damping.
• Treat single-site size, siting and penetration (11.5% to 34.5%) as limiting factors at the point of connection.

Who should read it

Grid planners and data-centre designers assessing the grid-stability impact and interconnection conditions of large AI data centres.

Document contents

1. Introduction
2. AI Workload Power Profile Modeling (A. Training Workload Modeling, B. Fine-tuning Workload Modeling, C. Aggregated AI Workloads for Datacenter Power Profiling)
3. Simulation Results (A. Simulation Settings, WECC 179-bus and NPCC 140-bus systems, eigen-analysis, penetration and siting sensitivity)
4. Oscillation Implications of AI Datacenter Deployment
5. Conclusion

Read the document

Publication details

Authors

Min-Seung Ko; Hao Zhu

Publisher

arXiv preprint (University of Texas at Austin; formatted for IEEE Transactions on Power Systems)

Year

2025

Document number

arXiv:2508.16457

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.48550/arXiv.2508.16457

Licence

CC BY 4.0 — share with attribution

Licence evidence

arxiv.org

Citation