⚡ electrical engineering

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

This paper proposes a decentralized switching-reference voltage control framework that leverages the structured, periodic nature of AI training workloads to effectively mitigate rapid voltage fluctuations in distribution systems while minimizing control costs and coordination requirements.

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

Published 2026-03-17

📖 4 min read☕ Coffee break read

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electrical grid as a massive, complex plumbing system. The "water pressure" in this system is voltage. If the pressure gets too high, pipes burst (equipment breaks); if it's too low, the water trickles out (lights dim, computers crash).

For decades, the system has been designed to handle steady, predictable water usage—like a family turning on a shower or a factory running a machine at a constant speed. The grid uses "pressure regulators" (voltage control) to keep things stable.

The New Problem: The "AI Gym"

Now, imagine a new type of building has moved into the neighborhood: a massive AI Training Data Center.

Think of this data center not as a steady factory, but as a gym full of synchronized weightlifters.

The Workout (Compute Phase): Every few seconds, thousands of weightlifters (AI chips) lift heavy weights simultaneously. This is a massive, sudden spike in power demand.
The Rest (Communication Phase): Then, they all stop lifting to catch their breath and talk to each other. The power demand drops just as suddenly.

This happens in a rhythmic, rapid cycle: LIFT! STOP! LIFT! STOP!

The Crisis: Traditional voltage regulators are like slow-moving garden hoses. They react to pressure changes after they happen. By the time the regulator tries to fix the pressure after the weightlifters drop their weights, the pressure has already swung wildly. To try to keep up, the regulator frantically opens and closes valves, wearing itself out and still failing to keep the pressure steady.

The Paper's Solution: "Smart Reference Switching"

The authors propose a new way to control the pressure. Instead of trying to force the pressure back to a single, rigid "perfect" level (like 1.0), they suggest moving the target.

Here is the analogy:
Imagine you are trying to catch a ball that bounces up and down between two heights.

Old Method (Fixed Reference): You stand still, waiting for the ball to come to your hand level. Every time the ball bounces up or down, you have to jump wildly to catch it. It's exhausting and you often miss.
New Method (Switching Reference): You realize the ball has a pattern. When the ball is high, you stand on a high stool. When the ball is low, you step down to a low stool. You move your "target height" to match the ball's rhythm. Now, you barely have to move your hands to catch it.

How It Works in the Real World

The paper introduces a decentralized system. This means every part of the grid (every neighborhood transformer) acts like a smart individual who doesn't need to call a central boss for permission.

Listening to the Rhythm: The local controller listens to the voltage. It sees the rapid "up and down" swings caused by the AI gym.
Calculating the Shift: It quickly figures out: "Okay, when the AI is working hard, the voltage naturally wants to drop. So, I will lower my 'target' voltage slightly to match that drop. When the AI rests, I'll raise my target back up."
The Result: Because the target moves with the load, the voltage never swings wildly. The "valves" (reactive power controls) don't have to work as hard. They stay calm and efficient.

Why This Matters

Less Wear and Tear: The grid equipment doesn't have to frantically adjust every second. It saves energy and extends the life of the hardware.
No Central Brain Needed: Each part of the grid figures this out on its own using local measurements. It's robust and fast.
Works with AI: Even if the AI data center tries to smooth out its own power usage (like the weightlifters trying to lift more slowly), this system adapts automatically. It doesn't need a complex handshake between the AI company and the power company.

In a Nutshell

The paper solves the problem of "fast, rhythmic power swings" from AI data centers by teaching the grid to dance with the load rather than fighting against it. Instead of trying to force a square peg into a round hole, they change the shape of the hole to fit the peg, keeping the system stable and efficient.

1. Problem Statement

The rapid deployment of large-scale Artificial Intelligence (AI) training workloads in hyperscale data centers introduces a new class of highly dynamic loads to power distribution systems. Unlike conventional loads that evolve gradually, AI training workloads exhibit rapid, periodic, and large-magnitude power fluctuations.

Root Cause: These fluctuations arise from the synchronized operation of GPU/TPU racks, which alternate between high-power compute phases (forward/backward propagation) and lower-power communication phases (synchronization/checkpointing).
Consequences: When connected to distribution feeders, these abrupt power swings cause significant voltage deviations, potentially exceeding allowable limits (e.g., $\pm5\%$ ).
Limitations of Existing Methods:
- Droop Control: Traditional decentralized volt-var droop control is designed for slowly varying loads. When applied to AI workloads, it reacts to every transition, leading to sustained oscillations, excessive reactive power control effort, and failure to mitigate worst-case voltage spikes effectively.
- Internal Coordination: While some approaches suggest internal data center load smoothing or coordination with external assets, these often require complex infrastructure or overlook the specific structured dynamics of AI training, leading to suboptimal performance.

2. Methodology

The authors propose a decentralized switching-reference voltage control framework that exploits the structured, two-mode behavior of AI training workloads rather than treating them as random noise.

A. System Modeling

Network Model: Uses the LinDistFlow model to relate bus voltages ( $v_t$ ) to active ( $p_t$ ) and reactive ( $q_t$ ) power injections: $v_t = R p_t + X q_t + \mathbf{1}$ .
Load Characterization: The data center load is modeled as switching between two modes ( $m_t \in \{0, 1\}$ $m_{t} \in {0, 1}$ ):
- Mode 0 (Communication): Lower power.
- Mode 1 (Computation): Higher power.
- Active power is modeled as $p_{j,t} = \bar{p}_j(m_t) + w_t$ , where $w_t$ represents intra-phase fluctuations.

B. Control Law Design

Instead of a fixed voltage reference (e.g., 1.0 p.u.), the proposed controller uses a time-varying reference $v^{ref}_t$ that switches between two levels aligned with the data center's operating modes.

Reference Structure: $v^{ref}_{i,t} = v^{bias}_{i,t} + s_{i,t} v^{amp}_{i,t}$ $v_{i, t}^{r e f} = v_{i, t}^{bia s} + s_{i, t} v_{i, t}^{am p}$ , where:
- $v^{bias}$ : The midpoint bias.
- $v^{amp}$ : The amplitude (half the voltage step induced by mode switching).
- $s_{i,t} \in \{-1, +1\}$ : The sign indicating the current mode.
Theoretical Basis: By aligning the reference voltage shift with the voltage change induced by the power step ( $\Delta v_{mode} = R(\bar{p}(0) - \bar{p}(1))$ ), the controller effectively "pre-emptively" absorbs part of the voltage swing. This reduces the disturbance term in the closed-loop error dynamics, theoretically halving the peak voltage deviation compared to fixed-reference control.

C. Measurement-Based Adaptation

Since the grid controller does not have direct access to the data center's internal mode state ( $m_t$ ) or optimal parameters, the reference is adapted solely based on local voltage measurements:

Amplitude Update ( $v^{amp}$ ): Estimated as half the maximum one-step voltage change observed in a recent time window, capturing the magnitude of the mode transition.
Bias Update ( $v^{bias}$ ): Calculated to center the two reference levels such that the steady-state voltage extrema are equidistant from the nominal 1.0 p.u. value, minimizing the worst-case deviation.
Sign Selection ( $s_{i,t}$ ): Determined by comparing the current local voltage to the estimated bias. If $v > v^{bias}$ , the system is in the high-power mode; otherwise, it is in the low-power mode.

3. Key Contributions

Novel Control Framework: Introduction of a decentralized switching-reference droop controller specifically tailored for the periodic, two-mode dynamics of AI training workloads.
Theoretical Guarantees: Establishment of conditions for voltage convergence and proof that the proposed switching reference minimizes the worst-case steady-state voltage deviation compared to fixed references.
Decentralized Implementation: Development of a fully local, measurement-based adaptation algorithm that requires no communication with the data center or knowledge of its internal state, ensuring scalability and compatibility with existing grid assets.
Compatibility: Demonstration that the method remains effective even when data centers employ internal load smoothing strategies, without requiring re-tuning or coordination.

4. Results

The proposed method was evaluated using simulations on the IEEE 33-bus distribution feeder with real-world power traces from a distributed AI training job (LLaMA-2-70B on H200 GPUs).

Single Data Center Scenario:
- Fixed-Reference Control: Resulted in persistent voltage oscillations and frequent saturation of reactive power control actions (hitting limits).
- Switching-Reference Control: Substantially reduced the magnitude and frequency of reactive power adjustments. Voltage remained consistently within the admissible 5% deviation band and closer to the nominal value.
Multiple Data Center Scenario:
- The controller successfully suppressed voltage deviations even when two data centers at different buses introduced simultaneous, asynchronous switching disturbances.
Compatibility with Load Smoothing:
- When internal load smoothing (using battery storage) was activated, the power fluctuations decreased. The proposed controller adapted automatically to the new power profile, maintaining effective voltage regulation without needing external coordination.

5. Significance

This work addresses a critical emerging challenge in power systems: the integration of massive, non-conventional loads driven by the AI boom.

Efficiency: It significantly reduces the "control effort" (reactive power cycling) required from inverters and storage systems, extending equipment lifespan and reducing operational costs.
Stability: It provides a robust mechanism to prevent voltage violations caused by the unique, high-frequency switching of AI workloads, which traditional control methods fail to handle.
Practicality: By relying only on local measurements and avoiding complex coordination with data centers, the solution is immediately deployable in existing distribution networks. It bridges the gap between the fast dynamics of AI computing and the slower dynamics of grid regulation.