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

arXiv:2508.16457 · eess.SY, cs.SY · Submitted 2025-08-22 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Wide-Area Power System Oscillations from Large-Scale AI Workloads".

Paper summary: Rosa: So, looking at this paper titled "Wide-Area Power System Oscillations from Large-Scale AI Workloads" by Min-Seung Ko and Hao Zhu, what does it really mean for us outside of the technical details?

Dev: It means we’re getting a way to proactively assess how deploying massive AI datacenters might stress the electrical grid through these sustained power fluctuations <ref:2508.16457#pg1>.

Taro: They are providing a tool for assessing the risk of inter-area and local oscillations that we might not see with traditional load models <ref:2508.16457#pg1>.

Rosa: The authors developed this new dynamic power profiling approach to model these AI workloads as stochastic, periodic profiles based on training and fine-tuning phases <ref:2508.16457#pg2>.

Dev: Their work points toward the need to improve future datacenter-dominated grid operations and how we integrate large-scale AI workloads into the power system <ref:2508.16457#pg2>.

Taro: It gives us a way to look at GenAI-dominant workloads and identify what those key factors are that shape the severity of these grid oscillations <ref:2508.16457#pg2>.

Rosa: The implication is that understanding how these workloads create persistent forcing inputs allows for better planning for grid stability when we scale up to a point where these facilities dominate electricity consumption <ref:2508.16457#pg3>.

Conclusion: Rosa: So, this paper looks at how big AI data centers can mess with the power grid, specifically looking at those wide-area oscillations they get when you have these massive workloads running continuously.

Dev: Right, it’s about modeling these loads not just as simple on and off switches or gradual ramps but as something more structured and continuous forcing a system.

Taro: They’re using a stochastic modeling framework to figure out the power consumption of training-focused AI datacenters by breaking down the cycle into compute and communication phases.

Rosa: So, the core idea here is that you can treat these AI workloads as periodic but random power consumption profiles based on how they divide their time between training and fine-tuning.

Dev: It lets them analyze how this structured forcing affects inter-area and local oscillations in the grid, which is different from just looking at traditional ramping models.

Taro: The methodology captures both the regular patterns of training cycles and the random fluctuations from those finer details of communication within each cycle.

Rosa: What this means for us is that we can start anticipating how scaling up AI power demands might push the grid toward instability through these specific types of power signals.

Dev: It gives us a way to understand what kind of persistent energy input from these facilities could trigger those oscillations in the power system.

Taro: The implication is that understanding the math behind these workloads helps us figure out how to keep the whole grid stable when AI consumption becomes a huge factor.

University of Texas at Austin

eess.SY, cs.SY

Submitted: 2025-08-22

Updated: 2026-04-20

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: This paper develops a new dynamic power profiling approach for modeling AI-centric datacenter loads and analyzing their impact on grid operations…

Key concepts

Dynamic Power Profiling
This is the new approach used to model AI datacenter power consumption. Instead of treating power as a simple ramping function, this method treats it as a continuous signal with structured patterns that change over time based on the specific AI tasks being performed.
Stochastic Modeling Framework
This is the mathematical tool developed to characterize how AI workloads consume power. It models these consumption profiles not just as fixed values but as random, periodic signals, accounting for the inherent randomness in compute and communication phases within each cycle.
Wide-Area Grid Oscillations
These are undesirable instability phenomena in large electrical grids where power flows between different regions cause waves or oscillations. The paper investigates whether the fluctuating power demands from AI datacenters can trigger these grid instabilities.

Terminology

Summary

This paper develops a new dynamic power profiling approach for modeling AI-centric datacenter loads and analyzing their impact on grid operations, particularly their potential to induce wide-area grid oscillations. p. 0

Where the paper says it

IEEE TRANSACTIONS ON POWER SYSTEMS 1 Wide-Area Power System Oscillations from Large-Scale AI Workloads Min-Seung Ko, Member, IEEE, and Hao Zhu, Senior Member, IEEE Abstract—This paper develops a new dynamic power profiling approach for modeling AI-centric datacenter loads and analyzing their impact on grid operations, particularly their potential to induce wide-area grid oscillations. We characterize the periodic stochastic power fluctuations inherent to large-scale AI workloads during both the training and fine-tuning stages, driven by the state-of-the-art graphics processing unit (GPU) computing architecture design.

Establishing AI datacenter workloads as continuous, structured forcing sources beyond the existing rampingcentric models allows for the analysis of their potential to induce inter-area and local oscillations. p. 1

Where the paper says it

By studying the oscillatory potentials in GenAI-dominant datacenter workloads, our work highlights an urgent need to enhance grid operations and stability amid the explosive growth in AI-driven demand. The main contributions of this work are summarized as follows: • Establishing AI datacenter workloads as continuous, structured forcing sources beyond the existing rampingcentric models, to allow the analysis of their potential to induce inter-area and local oscillations.

The work develops a stochastic modeling framework to characterize the power consumption of training-centric AI datacenters. p. 2
Where the paper says it

This concentration is reinforced by an architectural split, with centralized GPU clusters for training and finetuning, and inference handled in lower-power zones or edge facilities [19], [20]. STOCHASTIC POWER PROFILING FOR AI WORKLOADS We develop a stochastic modeling framework to characterize the power consumption of training-centric AI datacenters.

This framework models AI workloads as stochastic, periodic power consumption profiles by considering the compute- and communication-dominated phase division within each cycle. p. 1
Where the paper says it

Our goal is to enable an effective assessment of their potential to induce wide-area oscillations as the deployment of next-generation datacenters increases. We represent AI workloads with stochastic, periodic power consumption profiles, by considering the compute- and communication-dominated phase division within each cycle.

Modeling the Workload Signal

The stochastic modeling approach captures both periodic patterns and fine-scale randomness through distinct models for training and fine-tuning workloads.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Develop a stochastic AI workload model that captures both periodic patterns and fine-scale randomness to produce realistic, time-varying load profiles for grid studies (Section III). This model incorporates intra-iteration variability through Gaussian random processes for instantaneous deviations like eta up(t) and eta down(t).

  2. Enable effective assessment of potential wide-area oscillations by establishing AI workloads as continuous, structured forcing sources beyond the existing rampingcentric models to allow the analysis of their potential to induce inter-area and local oscillations (Contribution 1).

  3. Identify critical deployment factors for grid planning strategies, ranging from system inertia, datacenter penetration level, fluctuation frequency characteristics, size of datacenters, geographical dispersion to workload ratio (Contribution 4).

  4. Model the total fluctuating load demand as the superposition of stochastic power profiles from different workload types: P AI(t) = P tr0(t) + sum j P trj(t) + sum k P ftk(t) (Equation 11). This allows for explicit capture of large power fluctuation patterns at the second level.

  5. Inform grid planning by assessing how different factors shape oscillation severity, such as observing that narrower fluctuation bands, larger single-site capacities, or dispersed siting can intensify oscillations across multiple modes (Abstract).

Abstract

This paper develops a new dynamic power profiling approach for modeling AI-centric datacenter loads and analyzing their impact on grid operations, particularly their potential to induce wide-area grid oscillations. We characterize the periodic stochastic power fluctuations inherent to large-scale AI workloads during both the training and fine-tuning stages, driven by the state-of-the-art graphics processing unit (GPU) computing architecture design. % and distributed mini-batch processing cycles. 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. Using the WECC 179-bus system and the NPCC 140-bus system, we have numerically studied the amplitude and variability of oscillatory responses under different factors. These factors include system strength, penetration level, fluctuation frequency range, individual datacenter size, geographical deployment, fluctuation suppression level, and workload ratio. Simulation results show that, notably, narrower fluctuation bands, larger single-site capacities, or dispersed siting can intensify oscillations across multiple modes. Our models and numerical studies provide a quantitative basis for integrating AI-dominant electricity demand into grid oscillation studies and further support the development of new planning and operational measures to power the growth of AI/computing load demands.

Sources

Related papers