Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Switching-Reference Voltage Control for Distribution Systems with AI-Training Data Centers

Mingyuan Yan, Trager Joswig-Jones, Baosen Zhang, Yize Chen, Wenqi Cu

Abstract

Large-scale AI training workloads in modern data centers exhibit rapid and periodic power fluctuations, which may induce significant voltage deviations in power distribution systems. Existing voltage regulation methods, such as droop control, are primarily designed for slowly varying loads and may therefore be ineffective in mitigating these fast fluctuations. In addition, repeated control actions can incur substantial cost. To address this challenge, this paper proposes a decentralized switching-reference voltage control framework that exploits the structured behavior of AI training workloads. We establish conditions for voltage convergence and characterize an effective reference design that aligns with the two dominant operating levels of the AI training workload. The switching rule for voltage references is implemented solely using local voltage measurements, enabling simple local implementation while significantly reducing control effort. Simulation studies demonstrate that the proposed method substantially reduces both voltage deviations and reactive control effort, while remaining compatible with internal data center control strategies without requiring extensive coordination.

Switching-Reference Voltage Control for Distribution Systems with AI-Training Data Centers

Abstract

Large-scale AI training workloads in modern data centers exhibit rapid and periodic power fluctuations, which may induce significant voltage deviations in power distribution systems. Existing voltage regulation methods, such as droop control, are primarily designed for slowly varying loads and may therefore be ineffective in mitigating these fast fluctuations. In addition, repeated control actions can incur substantial cost. To address this challenge, this paper proposes a decentralized switching-reference voltage control framework that exploits the structured behavior of AI training workloads. We establish conditions for voltage convergence and characterize an effective reference design that aligns with the two dominant operating levels of the AI training workload. The switching rule for voltage references is implemented solely using local voltage measurements, enabling simple local implementation while significantly reducing control effort. Simulation studies demonstrate that the proposed method substantially reduces both voltage deviations and reactive control effort, while remaining compatible with internal data center control strategies without requiring extensive coordination.
Paper Structure (18 sections, 4 theorems, 25 equations, 5 figures)

This paper contains 18 sections, 4 theorems, 25 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Network Model and Problem Formulation
  3. Network Model and Data Center Power Characteristics
  4. Voltage Control and Challenges
  5. Grid-Side Voltage Control
  6. Control Law with Time-varying Voltage References
  7. Switching Law and Closed-Loop Voltage Deviation Analysis
  8. Measurement-Based Reference Adaptation
  9. Amplitude Update
  10. Bias Update
  11. Sign Selection
  12. Experiments
  13. Simulation Setup
  14. Decentralized Grid-Side Control
  15. Compatibility with Data Center Load Smoothing
  16. ...and 3 more sections

Key Result

Theorem 1

If the diagonal gain matrix $\mathbf K$ satisfies $0\prec \mathbf K \prec 2\mathbf X^{-1},$ then there exist a nonnegative matrix $\mathbf C$ and a constant $0<\epsilon<1$ such that for any initial time $t_0$ and all $t\ge t_0$, In particular, for any $\bar{\bm d}$ where $|\bm d_t|\le \bar{\bm d}$ for all $t$,

Figures (5)

  • Figure 1: Per unit power readings from an at-scale training job on DGX H100 racks choukse2025power.
  • Figure 2: Voltage response to step-like data center power transitions. (a) Active power profile at the data center bus, where the power switches between two operating levels and exhibits a step change of $\Delta p_{j,t_m}$ at each transition. (b) Voltage trajectory under decentralized droop control with a fixed reference $v^{\mathrm{ref}}\equiv 1$. Each transition induces a voltage shift approximately $R_{jj}\Delta p_{j,t_m}$. (c) Voltage trajectory with a time-varying reference aligned with the data center mode switching. By shifting the reference, the worst post-transition voltage deviation is reduced to approximately $\tfrac{1}{2} R_{jj}\Delta p_{j,t_m}$.
  • Figure 3: Single data center scenario: comparison of incremental reactive adjustment $\Delta\bm q$ (p.u.) and voltage trajectory $\bm v$ (p.u.) under droop control with a fixed reference and a switching reference. The proposed switching-reference control substantially reduces control effort in reactive power while maintaining voltage in admissible 5% voltage-deviation band.
  • Figure 4: Two data center scenario: comparison of incremental reactive adjustment $\Delta\bm q$ (p.u.) and voltage trajectory $\bm v$ (p.u.) under droop control with a fixed reference and a switching reference. The proposed switching-reference control remains effective even when multiple data center loads introduce simultaneous switching disturbances.
  • Figure 5: Decentralized grid-side voltage control with internal data center load smoothing. (a) Data center power profile with internal load smoothing starting at $t\approx1400$s. (b) Reactive control actions $\Delta\bm q$ (p.u.) and voltage trajectories $\bm v$ (p.u.) under the proposed switching-reference controller.

Theorems & Definitions (8)

  • Theorem 1: Convergence of voltage deviation
  • proof
  • Proposition 2: Voltage reference switching
  • proof
  • Proposition 3: Bias shifting
  • proof
  • Lemma 4
  • proof