Voltage Regulation in Distribution Systems with Data Center Loads
Yize Chen, Baosen Zhang · 2025
Read Voltage Regulation in Distribution Systems with Data Center Loads by Yize Chen, Baosen Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-06416.txt.
Introduction
Yize Chen and Baosen Zhang tap data halls them selves to calm the volt swings they cause, in a study from 2025. AI halls gulp power in bursts, and grids now feel more volt blips where they land. The halls, though, own a dial: chip clock and volt can dip for a spell with slight work loss.
Their scheme reads volt at each hall bus and trims feed by that dial, all local, no core call. The calm holds while the halls bear more load than a stiff scheme would dare. True LLM work traces drove the sims that prove it, and code plus data are free. The load that shakes the grid, they show, can still it.
Catalog record
| CREATOR | Yize Chen, Baosen Zhang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 24 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-06416
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Voltage Regulation in Distribution Systems with Data Center Loads Thanks: Y. Chen is partially supported by NSERC Discovery Grant RGPIN-2024-05015. B. Zhang is partially supported by the University of Washington Clean Energy Institute, the Grainger Foundation and the Galloway Foundation. Yize Chen, Baosen Zhang Abstract Recent boom in foundation models and AI computing have raised growing concerns on the power and energy trajectories of large-scale data centers. This paper focuses on the voltage issues caused by volatile and intensity of data center power demand, which also aligns with recent observations of more frequent voltage disturbances in power grids. To address these data center integration challenges, we propose a dynamic voltage control scheme by harnessing data center’s load regulation capabilities. By taking local voltage measurements and adjusting power injections at each data center buses through the dynamic voltage and frequency scaling (DVFS) scheme, we are able to maintain safe voltage magnitude in a distributed fashion with higher data center computing load. Simulations using real large language model (LLM) inference load validate the effectiveness of our proposed mechanism. Both the LLM power data and proposed control scheme are open sourced 1 1 1 https://github.com/chennnnnyize/voltage-regulation-with-data-centers . I Introduction Large-scale data centers are in the midst of a period of explosive growth driven by advances in AI and compute. Applications such as large language models (LLMs) and other foundational models such as GPT-4, Llama 3, and BERT have pushed the boundaries of modern AI while consuming an increasingly significant amount of power and energy [ 1 , 2 , 3 ] . The International Energy Agency (IEA) estimated that data centers consumed 460 TWh in 2022, accounting for 2% of global electricity consumption–a figure that is still rapidly increasing and expected to reach 5\% by 2030 [ 4 ] . Such energy burdens are generally observed in both the training and deployment stages of modern AI models [ 5 ] , and are further exacerbated by the growing cooling needs for more advanced GPU clusters [ 6 , 7 ] . Much attention has recently been paid to both the average and peak power demand of data centers. And for good reason. As electric power makes up more than 60% of the total operational cost of data centers [ 8 ] , they face the complex task of meeting customers’ quality of service requirements while balancing cost and sustainability. Meanwhile, power providers such as regional transmission operations and utility companies are having a challenging time coping with the ever-increasing and fluctuating data center demand. For example, despite skyrocketing electricity rates, California already has 270 data centers, and PG&E customers are planning 24 new large-scale data centers, with an average load forecast of 3.5 GW, equivalent to the power consumption of almost 3 million residential households [ 9 ] . Consequently, a number of works have proposed to reduce the total or peak power demand of data centers [ 10 , 11 , 5 ] . In this paper, we do not directly address either of these issues. Instead, we focus on voltage regulations in grids with data centers, an issue we believe to be critical to their integration into the power system but has perhaps received less attention. Data centers, as well as many other devices, are sensitive to the voltage magnitudes of their grid connections. A short voltage disturbance in Northern Virginia in July 2024 led to more than 60 hyperscale data centers disconnecting from the grid and almost caused widespread blackout in the PJM system [ 12 ] . A review by NERC also highlighted the challenges with voltage in data center load management [ 13 ] . Voltage regulation itself is a classical problem in power system operations. Most distribution systems have several types of equipment dedicated to voltage support, including tap-changing transformers, switched capacitors and STATCOMs. In systems with inverter-based resources, many algorithms have been proposed to use the power electronics in the inverters to inject reactive power for voltage support (see, e.g. [ 14 , 15 , 16 ] and the references therein). Fig. 1: Power consumption curve for one GPU during the model training of GPT2 124M. Multiple GPUs are ramping up and ramping down almost simultaneously, creating highly volatile and huge disturbances to distribution grid’s active power demand. Fig. 2: Power profiles of the four LLM inference tasks reveal distinct power ramping characteristics. Power is normalized in [0,1] range with thermal design power (TDP) limit marked. What makes data centers challenging and interesting in the context of voltage regulation is their unique load characteristics especially in the era of LLMs and foundational models [ 17 ] . Fig. 1 shows the single GPU’s power consumption profiles of training a GPT2 124M model on a cluster consisting of Nvidia A100 GPUs; each subfigure in Fig. 2 shows the power consumption profile of performing inference tasks using several common foundational models, and these inference timeseries is collected on a 8×A100 cluster instance. Also note that GPU power can be slightly over the Thermal design power (TDP) when GPU frequency is higher. We observe sharp dips in power consumption during both inference and training. 2 2 2 These dips are much more important in transient dynamics compared to more “steady-state” type of analysis. They are too short to have much impact on total energy consumption, and because they represent a decrease in power draw, they would not factor into peak demand at all. In these dips, a large down ramp is quickly followed (sometimes within a second) by a large up ramp, which can lead to large voltage deviations in the distribution system if left uncompensated. Because of the suddenness of the load changes, it is difficult to compensate for them using conventional devices such as tap-changing transformers or switched capacitors, which cannot act quickly enough to deal with (sub)second-level transients [ 14 ] . In addition, large data centers are often placed at the end of large lines, and since reactive power injections tend to be ”localized”, power electronic devices placed away from the data centers would have limited impact, despite their fast reaction capabilities. Therefore, in this paper, we analyze how data centers could adjust their active and reactive power injections to regulate voltage. For reactive power compensation, we adopt the standard technique in which we utilize power electronic rectifiers to change the amount of reactive power injected into the grid. To change the active power, we use dynamic voltage and frequency scaling (DVFS) to control the dynamic power draw of the GPUs. More specifically, we focus mostly on changing the frequency of the GPUs [ 18 , 19 ] . For example, NVIDIA GA100 supports operations at different frequencies in the range between 210 MHz and 1410 MHz [ 20 ] . Leveraging DVFS to adjust the underlying GPU’s frequency and power consumption in turn, it is possible achieve optimization at the iteration-level (spanning only milliseconds) opposed to the overall inference times (spanning several seconds). Since the power draw scales linearly with switching frequencies [ 21 ] , we can tradeoff compute with power. The circuit-level voltage (e.g., v_{dd} ) can also be changed to manage power consumption, but it appears to be more difficult to access,