pith. sign in

arxiv: 2604.18392 · v1 · submitted 2026-04-20 · 📡 eess.SY · cs.SY

Composite Control of Grid-Following Inverters for Stabilizing AI-Induced Fast Power Disturbances

Miroslav Kosanic , Marija Ilic This is my paper

Pith reviewed 2026-05-10 03:57 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords grid-following invertersdroop controlsingular perturbationAI data centerspower disturbancesstability analysiscomposite control
0
0 comments X

The pith

Singular perturbation analysis derives valid droop control for grid-following inverters to counter fast AI power transients.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that AI data centers produce rapid power changes that break the usual slow-fast separation in inverter controls. By applying singular perturbation techniques, the authors derive droop control laws from the requirement that the reduced slow system remains stable, rather than assuming the control form upfront. This approach also confirms that physical power delivery hardware filters AI loads into a class of disturbances with bounded rates of change. The resulting analysis provides explicit formulas for controller gains that guarantee disturbance rejection, along with conditions for the control to be physically realizable and for the maximum load ramp rate the system can handle.

Core claim

By modeling the inverter-AI load system with singular perturbation, physically-implementable droop control emerges as the stabilizing feedback for the reduced-order system, and AI workloads belong to a bounded-rate disturbance class due to hardware filtering, yielding gain bounds, a modulation admissibility condition, and a feasibility condition on maximum tolerable load ramp rate.

What carries the argument

Singular perturbation-based composite control that derives droop from reduced-system stability requirements.

If this is right

  • Droop gains can be tuned using explicit bounds tied to inverter parameters and disturbance rejection needs.
  • The control remains physically realizable only if the modulation admissibility condition holds.
  • Systems can tolerate AI load ramps up to a specific rate identified by the feasibility condition.
  • Grid-following inverters can stabilize without violating timescale separation assumptions through this derived control.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar techniques could apply to other fast renewable or storage controls facing rapid load variations.
  • Data center operators might use these bounds to size supplementary generation.
  • Extending to multi-inverter networks could reveal interaction effects not covered here.

Load-bearing premise

AI workloads produce power disturbances with rates bounded by physical filtering in power delivery hardware, allowing singular perturbation to apply despite traditional timescale separation being violated.

What would settle it

A simulation or experiment where the AI load ramp rate exceeds the derived feasibility condition and the system becomes unstable despite using the proposed control.

Figures

Figures reproduced from arXiv: 2604.18392 by Marija Ilic, Miroslav Kosanic.

Figure 1
Figure 1. Figure 1: Grid-following inverter with local AI load connected to stiff grid. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: visualizes the feasibility region for kd as a function of load ramp rate ρP for both the baseline design (V nom dc = 1200 V) and a high-voltage design (V nom dc = 1500 V). For each design, the lower bounds k SP d (timescale separation) and k ramp d (ramp tracking) define the minimum required gain, while k volt d (voltage saturation) defines the maximum. For the baseline design, the singular perturbation bo… view at source ↗
read the original abstract

AI data center loads create query-driven power transients on millisecond timescales. Such loads can violate the timescale separation assumptions underlying internal inverter control of grid-following resources collocated with data centers as supplementary generation. This paper develops a singular perturbation-based modeling and control for stabilizing fast power imbalances. We show that physically-implementable droop control is derived and valid by requiring reduced-system stability rather than being imposed a priori, and that AI workloads satisfy a bounded-rate disturbance class due to physical filtering in power delivery hardware. The analysis yields explicit gain bounds linking inverter parameters to disturbance rejection performance, a modulation admissibility condition ensuring physical realizability of the feedback linearizing control, and a feasibility condition identifying the maximum tolerable load ramp rate. Numerical simulations validate the theoretical predictions under stochastic AI transients.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript develops a singular perturbation-based composite control for grid-following inverters to stabilize AI-induced fast power disturbances on millisecond timescales. It derives droop control parameters by requiring stability of the reduced-order system rather than imposing them a priori. The key claim is that AI workloads belong to a bounded-rate disturbance class owing to physical filtering in power delivery hardware, which allows retaining singular perturbation analysis. This leads to explicit gain bounds linking inverter parameters to disturbance rejection, a modulation admissibility condition for physical realizability, and a feasibility condition on the maximum load ramp rate. The theoretical results are validated through numerical simulations with stochastic AI transients.

Significance. If the results hold, the work contributes a systematic way to handle load dynamics that violate traditional timescale separation assumptions in inverter control, by grounding the control design in reduced-system stability and providing concrete conditions for applicability. The explicit bounds and conditions are useful for practical implementation in data center collocated generation. The use of simulations for validation under realistic transients strengthens the practical relevance.

major comments (1)
  1. [Disturbance Modeling and Assumptions] The assertion that AI workloads satisfy a bounded-rate disturbance class due to physical filtering in power delivery hardware (as stated in the abstract and used to justify singular perturbation applicability) lacks supporting analysis or evidence. This assumption is critical because AI loads are described as violating traditional timescale separation; without establishing the rate bound independently (e.g., via hardware chain derivation or measurements), the reduced-system stability guarantee may not extend to the full-order system, undermining the derivation of the droop control and the explicit gain bounds.
minor comments (2)
  1. [Abstract] The abstract could more clearly distinguish between the derived droop control and the overall composite control strategy.
  2. [Notation] Some symbols for inverter parameters and disturbance rates could be defined more explicitly to aid readability.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and for highlighting the need to strengthen the disturbance modeling section. We address the major comment below and will incorporate revisions to improve the rigor of the bounded-rate assumption.

read point-by-point responses

  1. Referee: The assertion that AI workloads satisfy a bounded-rate disturbance class due to physical filtering in power delivery hardware (as stated in the abstract and used to justify singular perturbation applicability) lacks supporting analysis or evidence. This assumption is critical because AI loads are described as violating traditional timescale separation; without establishing the rate bound independently (e.g., via hardware chain derivation or measurements), the reduced-system stability guarantee may not extend to the full-order system, undermining the derivation of the droop control and the explicit gain bounds.

    Authors: We agree that the manuscript would benefit from a more explicit derivation of the bounded-rate disturbance class. The current text invokes physical filtering in power delivery hardware (e.g., server PSUs and associated cabling) as the mechanism that enforces the rate bound, but does not provide a dedicated derivation or supporting references. In the revised version we will add a short subsection (likely in Section II) that derives an explicit upper bound on the disturbance rate from typical hardware time constants. This will include (i) a first-order low-pass model of PSU dynamics with measured time constants from data-center literature, (ii) a simple cascade analysis showing that the effective power ramp rate seen by the inverter is bounded by the slowest filter pole, and (iii) a brief discussion of how this bound remains valid even under stochastic AI query arrivals. The added material will directly support the applicability of singular perturbation analysis and the subsequent gain bounds without altering the core theoretical results or simulation validation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation from stability requirement is independent

full rationale

The paper derives physically-implementable droop control by imposing stability on the reduced-order system via singular perturbation analysis, rather than imposing it a priori or fitting parameters to match a target outcome. This follows standard control-theoretic reduction and does not reduce to the inputs by construction. The bounded-rate disturbance class for AI loads is asserted as a consequence of physical hardware filtering to justify retaining the singular perturbation framework, but this is presented as an enabling assumption without evidence that it is constructed from or equivalent to the derived droop law or gain bounds. No self-citations, ansatzes smuggled via prior work, or renamings of known results are load-bearing in the abstract or described chain. The explicit gain bounds, modulation admissibility, and ramp-rate feasibility conditions follow from the stability analysis on the reduced system. The paper is self-contained against external benchmarks in its modeling approach.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard singular perturbation assumptions for timescale separation in inverter dynamics and a domain-specific claim that physical hardware filters AI power demands into bounded-rate disturbances. No free parameters or invented entities are identified in the abstract.

axioms (2)
  • domain assumption Singular perturbation theory applies to separate fast and slow dynamics in the grid-following inverter system under AI load transients.
    Invoked to develop the composite control and reduced-system stability analysis.
  • domain assumption AI workloads produce power disturbances whose rate of change is bounded due to physical filtering in power delivery hardware.
    Used to define the disturbance class for which gain bounds and feasibility conditions are derived.

pith-pipeline@v0.9.0 · 5431 in / 1496 out tokens · 71439 ms · 2026-05-10T03:57:57.382240+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages

  1. [1]

    The unseen AI dis- ruptions for power grids: LLM-induced transients,

    The unseen AI disruptions for power grids: LLM-induced transients , author=. arXiv preprint arXiv:2409.11416 , year=

    work page arXiv

  2. [2]

    Empirical measurements of AI training power demand on a GPU-accelerated node,

    Empirical measurements of AI training power demand on a GPU-accelerated node , author=. arXiv preprint arXiv:2412.08602 , year=

    work page arXiv

  3. [3]

    AI load dynamics–a power electronics perspective,

    AI Load Dynamics--A Power Electronics Perspective , author=. arXiv preprint arXiv:2502.01647 , year=

    work page arXiv

  4. [4]

    IEEE Transactions on Power Systems , volume=

    Stability analysis of distributed power generation with droop inverters , author=. IEEE Transactions on Power Systems , volume=. 2015 , publisher=

    work page 2015

  5. [5]

    IEEE Transactions on Smart Grid , volume=

    Dynamic phasors-based modeling and stability analysis of droop-controlled inverters for microgrid applications , author=. IEEE Transactions on Smart Grid , volume=. 2014 , publisher=

    work page 2014

  6. [6]

    Energy Markets and Responsive Grids: Modeling, Control, and Optimization , pages=

    A hierarchy of models for inverter-based microgrids , author=. Energy Markets and Responsive Grids: Modeling, Control, and Optimization , pages=. 2018 , publisher=

    work page 2018

  7. [7]

    2022 58th Annual Allerton Conference on Communication, Control, and Computing (Allerton) , pages=

    Dynamic performance of unified grid-forming/following inverter control , author=. 2022 58th Annual Allerton Conference on Communication, Control, and Computing (Allerton) , pages=. 2022 , organization=

    work page 2022

  8. [8]

    IEEE transactions on power systems , volume=

    Nonlinear dynamic load models with recovery for voltage stability studies , author=. IEEE transactions on power systems , volume=. 2002 , publisher=

    work page 2002

  9. [9]

    Ajala and S

    O. Ajala and S. V. Dhople and F. Milano , title =. Electric Power Systems Research / Lecture Notes in Electrical Engineering , year =

    work page

  10. [10]

    Caduff and others , title =

    I. Caduff and others , title =. Electric Power Systems Research , year =

    work page

  11. [11]

    SIAM Journal on Control and Optimization , volume=

    Near--Optimal Feedback Stabilization of a Class of Nonlinear Singularly Perturbed Systems , author=. SIAM Journal on Control and Optimization , volume=. 1978 , publisher=

    work page 1978

  12. [12]

    Lin and others , title =

    Y. Lin and others , title =

    work page

  13. [13]

    Jafarpour and F

    S. Jafarpour and F. D. Singular Perturbation and Small-signal Stability for Inverter Networks , journal =

    work page

  14. [14]

    Wu and others , title =

    J. Wu and others , title =. Frontiers in Energy Research , year =

    work page