PaperlayerView exact revision on arXiv

Source-linked AI summary

Designing Grid-Aware Dynamic Specifications for Large Data Center Loads

Ashutossh Gupta, Vassilis Kekatos

arXiv:2609.18888v1eess.SYcs.SY

TL;DR

Large data center loads can produce abrupt ramps and sustained oscillations, yet utilities’ dynamic limits are sparse and inconsistent. The paper derives ramping and spectral specifications for LLM training loads, using COI frequency analysis and Fourier-coefficient constraints. It reports nonlinear WECC validation of the ramping approximation and a compact ellipsoidal representation of admissible oscillatory-load coefficients.

  • Problem

    Large data centers’ abrupt ramps and sustained oscillations can challenge grid frequency dynamics, while existing utility specifications are sparse and inconsistent.

    Source: §I · p. 1

  • Method

    The paper derives ramping specifications from COI frequency nadir and RoCoF analysis, and spectral specifications from constraints on Fourier-series coefficients.

    Source: §I · p. 2

  • Results

    Nonlinear simulations of the full 179-bus WECC system show ramp-load nadir and RoCoF are well approximated by COI frequency; oscillatory-load admissible coefficients have a tight inner ellipsoidal approximation.

    Source: §VII · p. 9

  • Takeaways & Limitations

    The framework specifies ramp times, load amplitudes, and oscillatory Fourier coefficients to keep grid-frequency dynamics within acceptable limits.

    Source: §I · p. 2, §VII · p. 9

Abstract

from arXiv · show

As data center (DC) loads increasingly penetrate the power grid, there is an urgent need for grid operators to provide clear dynamic specifications to DC owners to ensure safe grid operation. To this end, we study two salient behaviors of large language model (LLM) training loads: abrupt ramps at job initiation and termination, which induce transient frequency excursions, and sustained periodic oscillations during training, which result in oscillatory steady-state behavior. For ramping loads, we show that nodal rotor frequencies can be accurately approximated by the center-of-inertia (COI) frequency and derive analytical expressions for its nadir and rate of change of frequency (RoCoF). These expressions determine allowable combinations of ramp times and steady-state load demands satisfying prescribed frequency limits. For oscillatory loads, we derive spectral specifications on their Fourier coefficients and show that the admissible coefficient set can be approximated by a polytope. We further obtain a compact representation via its maximum-volume inscribed ellipsoid, which we show is axis-aligned. Numerical tests on the WECC 179-bus system demonstrate that the resulting specifications remain valid for higher-order nonlinear dynamics. The proposed framework provides actionable specifications for regulating the dynamic behavior of large DC loads and informing load-shaping mechanisms within the data center ecosystem.

I. INTRODUCTION

Large data centers’ steep ramps and sustained power fluctuations pose grid-stability challenges, while existing dynamic limits are sparse and inconsistent. The paper develops ramping and spectral specifications for LLM training loads and tests the resulting analyses on nonlinear WECC dynamics.

I. INTRODUCTION

  • Sparse and inconsistent utility limits leave a need for uniform interconnection-level dynamic specifications for large loads.Existing approaches such as workload shifting do not curb grid oscillations at fine temporal scales.Source: §I · p. 1
  • LLM training creates two grid-relevant load behaviors: abrupt job-start/stop ramps and sustained periodic fluctuations during training.These behaviors can cause frequency excursions or excite low-frequency oscillations when their frequency content aligns with lightly damped modes.Source: §I · pp. 1–2
  • The paper formulates ramping and spectral specifications to keep frequency nadir and RoCoF within prescribed limits.For ramps, nodal rotor frequencies are approximated by the center-of-inertia frequency, yielding analytical nadir and RoCoF expressions for allowable ramp-time and load-demand combinations.Source: §I · p. 2
  • For sustained oscillatory loads, admissible Fourier coefficients are approximated by a polytope whose maximum-volume inscribed ellipsoid is axis-aligned.The paper reports that numerical tests using ANDES validate the analytical results under higher-order nonlinear generator dynamics.Source: §I · p. 2

II. PROBLEM STATEMENT

LLM training loads combine rapid power changes at job start and completion with periodic demand fluctuations during training. This section frames separate ramping and spectral design problems around keeping grid-frequency nadir and RoCoF within limits.

II. PROBLEM STATEMENT

  • Milliseconds-scale transitions between idle and full utilization, together with periodic GPU computation and communication phases, create transient and oscillatory load behavior.The oscillation’s fundamental frequency is typically f0 ∈[0.5, 1.5] Hz.Source: §II · p. 2
  • Ramping specifications choose allowable ramp time T and steady-state demand P so grid frequency remains within safe operating limits.The stated steady-state demand range is P ∈[50, 500] MW.Source: §II · p. 2
  • Spectral specifications constrain Fourier-series coefficients of the smoothed periodic load to keep load-induced frequency dynamics benign.The constant coefficient c0 models average training power, while the nonzero coefficients describe oscillatory components that could excite grid-frequency oscillations.Source: §II · pp. 2–3
  • Both design problems use frequency nadir and RoCoF as performance metrics, with stricter limits proposed for normal operation than emergency-event limits.The stated reason is that LLM-induced dynamics occur during normal grid operation.Source: §II · p. 3

III. POWER SYSTEM DYNAMICS MODELING

The section models how power injections affect generator frequencies, then uses simplifying assumptions to decompose the grid response into single-input, single-output eigensystems. The first eigensystem yields the common center-of-inertia frequency, while the remaining eigensystems capture bus-specific deviations.

III. POWER SYSTEM DYNAMICS MODELING

  • The model maps power injections to generator-frequency deviations using a MIMO system, with inertia, damping, turbine, and inverse-droop parameters represented in diagonal matrices.The bus-coupling matrix is the negative Jacobian of nodal active power injections with respect to voltage angles and is often approximated by the Kron-reduced bus susceptance matrix.Source: §III · p. 3
  • The MIMO grid-frequency model can be decomposed into N single-input, single-output eigensystems under assumptions that damping and inverse droop scale with inertia and turbine constants are identical.The assumptions enable analytical insight, and the proposed specifications are validated against realistic dynamic grid models.Source: §III · p. 3
  • The first eigensystem responds to aggregate system input and supplies the common center-of-inertia frequency; the remaining eigensystems capture deviations from it.The response is represented as a linear combination of the eigensystems.Source: §III · p. 4
  • The WECC 179-bus benchmark is Kron-reduced to a 29-machine network, and its parameters do not exactly satisfy the simplifying assumptions.The specifications are evaluated on a detailed nonlinear WECC model.Source: §III · p. 4
  • The WECC eigensystems act as band-pass filters with approximately constant bandwidth and gain, while their center frequencies increase with eigensystem index.Figure 2 shows frequency responses for the first ten eigensystems.Source: §III · p. 4, Figure 2 · p. 4

IV. RAMPING SPECIFICATIONS FOR LARGE LOADS

The section derives frequency nadir and RoCoF limits for ramping loads from the COI response, approximating all bus frequencies by that response. The resulting specifications identify allowable combinations of ramp time and steady-state load demand.

IV. RAMPING SPECIFICATIONS FOR LARGE LOADS

  • For a 594 MW ramp with T = 6 s, the maximum frequency deviation is 0.025 Hz, with no oscillations and bus frequencies moving together around the COI.In the generator-trip comparison, frequencies deviate considerably from the COI and oscillate with a 0.05 Hz peak-to-peak amplitude.Source: Figure 3 · p. 5, §IV · p. 5
  • The ramp-load test supports approximating bus rotor frequencies by the COI frequency, unlike the generator-trip test, where bus frequencies deviate from the COI.The ramp input's spectrum has first-order decay near zero frequency and second-order decay at higher frequencies, motivating focus on the first eigensystem for ramping loads.Source: §IV · pp. 4–5, Figure 3 · p. 5
  • Both nadir and RoCoF scale linearly with load amplitude P, while their dependence on T is more subtle and can be tabulated to identify allowable pairs.For the tested inputs, the nadir occurs after T and RoCoF before T; both decrease as T increases.Source: §IV · pp. 5–6
  • The COI nadir is a non-monotonic function of ramp time; when the first-eigensystem step response is negative through T, the nadir occurs after the ramp ends.Equation (21) defines a feasible region in (T, P) space for a prescribed nadir limit.Source: §IV · p. 6
  • The maximum COI RoCoF is the larger of two candidate values, and rτ < 1 is a sufficient condition for it to occur before T.The condition is sufficient, not necessary: WECC does not satisfy it, yet its RoCoF still occurs before T.Source: §IV · p. 6
  • Given prescribed nadir and RoCoF limits, operators can use the derived expressions to specify allowable ramp-time and load-demand pairs, independent of load location.Numerical tests corroborate the derived results for actual bus frequencies and examine load-location effects.Source: §IV · p. 6

V. SPECTRAL SPECIFICATIONS FOR PERIODIC LOADS

The section specifies allowable Fourier-series coefficients for periodic loads so induced rotor frequencies satisfy prescribed nadir and RoCoF limits. It approximates the coefficient constraints by a polytope and offers a compact, axis-aligned ellipsoid representation.

V. SPECTRAL SPECIFICATIONS FOR PERIODIC LOADS

  • The admissible set of nonzero Fourier-series coefficients is designed to keep induced rotor frequencies within prescribed nadir and RoCoF limits, excluding the average-load term c0 = P.The design requires the frequency response between source and destination buses, which can be computed, simulated, or experimentally estimated.Source: §V · pp. 6–7
  • Truncating the Fourier expansion to K coefficients and enforcing constraints at discrete samples converts the infinite constraint set into finitely many linear inequalities.Sampling at fs ≫ 2Kf0 and over one fundamental period yields L = fs/f0 samples.Source: §V · p. 7
  • The resulting allowable coefficient vectors form a polytope, providing a feasible-set representation for periodic-load specifications.The frequency deviations and RoCoF samples are linear functions of the coefficients.Source: §V · p. 7
  • The ellipsoid offers a more compact specification: describing it requires communicating a single complex-valued vector rather than the many vectors needed for the polytope.The maximum-volume ellipsoid can be found with a convex optimization problem.Source: §V · p. 7
  • The maximum-volume ellipsoid inscribed in the polytope is axis-aligned because discrete-time shifts of the periodic waveform preserve admissibility.This symmetry makes the ellipsoid's shape matrix diagonal.Source: §V · p. 7

VI. NUMERICAL TESTS

The tests evaluate ramping and spectral specifications on the WECC 179-bus system, including validation against detailed nonlinear dynamics. Ramping constraints depend on ramp speed, while admissible oscillatory-load content varies with harmonic frequency, bus location, and loading conditions.

VI. NUMERICAL TESTS

  • The ramping tests use a 0.03 Hz nadir limit and a 0.02 Hz/s RoCoF limit on a 29-generator WECC model with 60.78 GW base load.The model is used as a robustness test because it does not satisfy both simplifying assumptions.Source: §VI · p. 7
  • For T < 1.5 s, RoCoF binds the admissible (T, P) values; for slower ramps, the nadir constraint binds.Fig. 5 identifies parameter combinations satisfying both limits for the WECC 179-bus system.Source: §VI · p. 8, Figure 5 · p. 8
  • For T > 3 s, the COI-based design predicts nadir and RoCoF accurately, but under-approximates both for faster ramps.The comparison uses maximum nadir and RoCoF across generator frequencies in nonlinear simulations.Source: §VI · p. 8, Figure 6 · p. 8
  • Nearly all coefficient vectors inside the blue region satisfy the RoCoF limit, while most outside violate it, validating the ellipsoidal boundary against nonlinear dynamics.The test uses a linearized-model specification and a detailed nonlinear model.Source: §VI · p. 8, Figure 7 · p. 8
  • Bus 50 tolerates considerably less harmonic content than the other tested buses, and scaling loads by ±5% changes the allowable βn.The tests vary fundamental frequency from 0.3 to 1 Hz and examine buses 35, 50, 75, and 138.Source: §VI · pp. 8–9, Figure 8 · p. 9

VII. CONCLUSIONS

The paper presents dynamic requirements for data center loads based on ramping and oscillatory behavior. WECC nonlinear simulations support the ramping approximation, while the inscribed ellipsoid provides a tight inner approximation for admissible oscillatory-load coefficients.

VII. CONCLUSIONS

  • WECC nonlinear simulations show that COI frequency well approximates ramp-induced nadir and RoCoF, while an inscribed ellipsoid tightly approximates admissible Fourier-series coefficients.The frequency response depends on injection location, so buses tolerate different harmonic content.Source: §VII · p. 9

APPENDIX

The appendix derives nadir and RoCoF conditions from stationary points of the COI response. It also uses phase-rotation symmetry of feasible Fourier coefficients to establish that the maximum-volume inscribed ellipsoid is axis-aligned.

APPENDIX

  • The nadir proof locates extrema through zeros of the COI response derivative and selects the earliest minimum after T.A negative step response through T makes the COI frequency decrease monotonically over that interval; the exponential envelope then identifies the earliest post-T minimum.Source: APPENDIX · p. 9
  • The appendix identifies ωs(t) as the first eigensystem’s step response and uses its first local maximum to test whether all local maxima are negative.The decaying exponential term makes the first maximum the greatest.Source: APPENDIX · p. 9
  • The RoCoF proof finds stationary points before and after T, with the decaying exponential making the first post-zero stationary point attain the maximum magnitude.Comparing the two intervals yields a sufficient condition c < 2ab for the maximum to occur during (0, T).Source: APPENDIX · pp. 9–10
  • Shifting a periodic response by ℓ′/fs rotates each Fourier harmonic’s phase while preserving whether the nadir and RoCoF constraints hold.The phase rotations are collected in the diagonal matrix G(ℓ′), which maps feasible coefficients to feasible coefficients.Source: APPENDIX · p. 10
  • Rotational symmetry and uniqueness of the maximum-volume inscribed ellipsoid imply that its shape matrix is diagonal.The proof shows rotated ellipsoids retain the same volume and remain inscribed in the feasible coefficient set.Source: APPENDIX · p. 10
Loading 2609.18888v1…