Pith. sign in

REVIEW 3 major objections 4 minor 64 references

The paper claims that islanded AI data centers can run reliably on gas turbines plus grid-forming batteries during the years before grid interconnection, holding frequency and voltage within tight limits.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 18:06 UTC pith:UHSCXTDD

load-bearing objection Useful phased-islanding framework for AI data centers, but the 'reliably support for months' claim lacks an energy budget; worth a serious referee. the 3 major comments →

arxiv 2607.17391 v1 pith:UHSCXTDD submitted 2026-07-19 eess.SY cs.AIcs.SY

A Phased Development Framework Enabling Islanded Operation of Sustainable AI Data Centers With Onsite Grid-Following and Grid-Forming Energy Architectures

Soham Ghosh , Nabil Mohammed , Mohammad Ashraf Hossain Sadi This is my paper
classification eess.SY cs.AIcs.SY
keywords AI data centerislanded operationgrid-forming inverterbattery energy storagegas turbinephased deploymentfrequency regulationelectromagnetic transient simulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes a phased construction framework that lets hyperscale AI training data centers operate on on-site generation for months or years while grid interconnection is still years away. It argues that a hybrid of natural gas turbines and grid-forming battery energy storage can reliably serve a 300 MW AI training load, keeping voltage within 0.95-1.05 per unit and frequency within 59.9-60.25 Hz. Electromagnetic transient simulations across seven cases show that gas turbines alone cannot handle AI load spikes, but adding a grid-forming BESS stabilizes the island. Once the grid connection matures, the same hybrid supports grid-connected operation and can re-island during grid disturbances. The paper also compares grid-forming control strategies and recommends an adaptive virtual synchronous generator for smooth reconnection and restoration.

Core claim

This study demonstrates through EMT simulations that a 300 MW AI training data center can be powered in islanded mode by a combination of on-site gas turbines and a grid-forming battery energy storage system (GFM BESS), and in some configurations by a GFM BESS alone. The GFM BESS provides the voltage and frequency reference, absorbing the rapid power swings characteristic of GPU training loads while the slower gas turbines supply base power. Across Cases 3, 4, and 5, the system holds MV bus voltage within 0.95-1.05 p.u. and frequency within 59.9-60.25 Hz. Case 6 shows that after full grid interconnection, the same on-site system can seamlessly pick up the load when the grid disconnects. Case

What carries the argument

The central mechanism is the hybrid islanded microgrid formed by a grid-forming battery energy storage system (GFM BESS) that establishes the voltage and frequency reference at the 13.2 kV MV bus, working alongside slower on-site gas turbines and grid-following inverters. The GFM BESS uses droop-based P-omega and Q-V control to rapidly inject or absorb active power, absorbing the high-frequency, low-magnitude swings of AI training loads. The load model is a single NVIDIA GB200 GPU trace scaled to represent a 300 MW plant; this scaling is the load-bearing input for all simulation cases.

Load-bearing premise

The simulations assume that a real 300 MW AI training facility's power draw behaves like a single scaled GPU trace, so if actual aggregated load swings are larger or more jagged, the claimed voltage and frequency bands may not hold.

What would settle it

Measure the aggregate power draw of a large AI training cluster (e.g., a few hundred MW) over a training run and compare the observed ramp rates and frequency/voltage deviations against Fig. 8a and the Case 3/4 simulation results; if deviations exceed 0.05 p.u. voltage or 0.35 Hz frequency, the central claim collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Data center developers could begin operating a 300 MW AI training facility within about 24 months using on-site gas turbines and GFM BESS, without waiting for interconnection studies.
  • A GFM BESS-only configuration (Case 4) offers a path for sites where air-quality permitting rules out gas turbines, with the BESS alone maintaining tight voltage and frequency bands.
  • The same hybrid architecture supports the transition to full grid interconnection and can re-island automatically during grid disturbances, providing ride-through without dropping the AI load.
  • The comparison of grid-forming controls indicates that adaptive VSG control is preferable for reconnecting and restoring AI data center loads to the grid, especially under varying grid strengths.
  • The framework implies that load ramp-rate limits and large-load model validation standards should be standardized across jurisdictions to prevent simultaneous disconnections of large AI loads.

Where Pith is reading between the lines

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

  • If the scaled load profile accurately represents real hyperscale training, this framework suggests that early AI deployments may not need to fund long-lead transmission upgrades, freeing capital for compute capacity.
  • The paper's reliance on a single-tenant, non-diversified load profile is conservative, but real multi-tenant facilities could have different swing shapes; the framework would be stronger with aggregated load telemetry from multiple sites.
  • A natural extension is to couple this architecture with GPU power-smoothing software (e.g., setting ramp-rate floors) to reduce BESS sizing requirements, a combination the paper mentions but does not simulate.
  • The recommended adaptive VSG control for restoration implies that control schemes should be grid-strength aware; a testable next step is hardware-in-the-loop validation of the adaptive VSG on a scaled data center testbed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. This paper proposes a phased engineering, procurement, and construction framework that enables a 300 MW AI training data center to operate islanded during the approximately 24–36-month period before grid interconnection, using on-site natural gas turbines and grid-forming (GFM) battery energy storage — with a nominally included PV component — and then to connect to the grid and re-island on grid disturbances. The authors develop an EMT simulation model of the 13.2 kV MV bus and test seven cases: gas turbines alone, gas turbines with grid-following BESS-PV, gas turbines with GFM BESS-PV, GFM BESS-PV alone, GFM+GFL BESS-PV, a grid-connected-to-islanded transition, and a comparison of GFM control strategies for restoration. The central claims are that gas turbines alone cannot track AI training transients (frequency excursion to 61.5 Hz in Case 1), while a hybrid of GFM BESS and gas turbines can maintain MV bus voltage within 0.95–1.05 p.u. and frequency within 59.9–60.25 Hz (Cases 3–5), and that adaptive VSG control is the most suitable GFM strategy for reconnection. The AI training load profile used in Cases 1–6 is obtained by linearly scaling a single NVIDIA GB200 GPU trace to 300 MW.

Significance. The paper addresses a timely and practically important problem: the mismatch between hyperscale AI data center deployment schedules and grid interconnection timelines. If the quantitative claims are validated, the proposed phased architecture — gas turbines plus GFM BESS during islanded early/intermediate phases, transitioning to grid-connected operation with islanding capability — would offer a concrete path for developers. The paper has clear strengths: a well-structured case taxonomy with explicit recommended and non-recommended configurations, realistic procurement lead-time tables, and a transparent statement that single-tenant load aggregation is conservative. The qualitative conclusion that gas turbines alone are ill-suited to AI training transients while GFM BESS can provide fast power tracking is consistent with established converter-control behavior. However, the central 'reliably support data center operations' claim is not yet quantitatively supported because (i) BESS systems are specified only in MVA with no energy capacity or state-of-charge modeling, and (ii) the 300 MW aggregate load is a scaled single-GPU trace with no validation against measured hyperscale data. C

major comments (3)
  1. [§V, Table VIII, Cases 3–5] The central claim that the proposed architectures 'reliably support data center operations' during the stated 24–36-month islanded phases is not supported by the simulations as presented. Every BESS is specified only in MVA (200 MVA Case 3, 350 MVA Case 4, 300 MVA Case 5); no MWh rating, state-of-charge model, or PV generation profile is given. In Case 4, a 350 MVA GFM BESS alone serves a 300 MW load with no gas turbines; for any finite energy capacity, the battery will deplete in hours, far short of months, unless unmodeled PV supplies continuous net energy. In Case 3, 4×40 MVA gas turbines (≈160 MW) cannot cover even the average of the 200–300 MW load, so the BESS must supply net energy over long intervals. The EMT plots appear to cover seconds and never show SOC or energy throughput. A SOC-constrained simulation, or at minimum an energy-budget calculation with explicit MWh ratings and
  2. [§V opening, Fig. 4, §V-A, §VI-C] The 300 MW plant load is obtained by linearly scaling a single NVIDIA GB200 GPU training trace to plant scale. This is the input to all Cases 1–6 and directly determines the reported voltage and frequency excursions, including the 61.5 Hz excursion in Case 1 and the 59.9–60.25 Hz band in Case 4. No validation against measured hyperscale facility data is provided. The paper itself notes that single-tenant operation is 'conservative' (§V-A) and recommends anchoring plant models to measured data (§VI-C), but it does not test sensitivity to alternative aggregation, diversification, or GPU power-smoothing assumptions. Without such analysis, the quantitative thresholds are not robust to the central modeling assumption.
  3. [§V-G, Fig. 13, Table IX, Refs [59]–[60]] The conclusion that adaptive VSG control is the most suitable GFM strategy for AI data center restoration is not demonstrated by new simulations in this manuscript. Figure 13 is reproduced from the authors' prior work [59], and Table IX is a qualitative scoring table; the text refers the reader to [59], [60] for details. This makes Case 7 a literature-based comparison rather than an EMT result of the present study. Please either include the supporting time-domain simulations and parameter values, or explicitly state that the Case 7 conclusion is drawn from previous publications and is not newly demonstrated here.
minor comments (4)
  1. [§V-D, Fig. 12] Typos: 'Fig. Fig. 10a' should be 'Fig. 10a'; 'upto' should be 'up to'. Also 'SMEs' appears in §III-C4 for 'SMES'.
  2. [Table VII, Case 6 row] The phrase 'Final built represents a higher percent of cleaner GFM BESS-PV system' is awkward; consider 'final configuration' and 'larger share'.
  3. [Throughout, Table VIII] The term 'BESS-PV' implies PV generation, but no PV capacity, irradiance, or time-of-day profile is given. Clarify whether the PV is included in the stated MVA ratings or is purely nominal; otherwise the name is misleading.
  4. [§V-C] The voltage threshold is described as '0.95–1.05 p.u.' in Case 3, while Case 4 calls 13.1–13.5 kV 'tight margins'. State whether these are equipment limits, design targets, or regulatory thresholds, and use consistent language.

Circularity Check

1 steps flagged

Central islanding simulations are self-contained and non-circular; the circularity-adjacent burden is in Case 7, where the GFM-control comparison and adaptive-VSG recommendation are imported from the authors' own prior papers.

specific steps
  1. self citation load bearing [§V-G (Case 7), Fig. 13 caption, Table IX, Conclusion]
    "Fig. 13: Performance of the GFMIs in a strong grid with SCR = 8 and Xg/Rg = 7 ... [59]. ... For a more detailed characterization of the underlying GFM techniques, including both frequency-domain and time-domain analyses, the interested reader is referred to the comprehensive studies reported in [59], [60]. ... an adaptive virtual synchronous generator–based control scheme is recommended because of its reduced sensitivity to short-circuit strength variations and its improved damping of transient oscillations under changing grid conditions."

    The manuscript's contribution (3) claims that 'different GFM control strategies are assessed in terms of small-signal stability, transient performance, and voltage–frequency support capabilities for AI data center applications.' In Case 7, however, the substantive comparison is not derived in this paper: Fig. 13 is explicitly attributed to [59], the surrounding text defers to [59], [60] for the underlying analyses, and the final A-VSG recommendation is attributed to [64]. [59], [60], and [64] are all authored or co-authored by N. Mohammed, a co-author of this manuscript. Thus the GFM-control comparison and the adaptive-VSG recommendation reduce to a self-citation chain: the new 'AI data center restoration' framing is added, but no new simulation, derivation, or verification of the comparat

full rationale

The primary islanded-operation claims (Cases 1–6) are not circular: they are EMT simulations with explicit machine and inverter models (Section IV) and stated parameters (Table VIII), and the reported voltage/frequency metrics are simulation outputs, not inputs. The paper's scaling of a single-GPU AI training trace to a 300 MW plant is an assumption about the load scenario, not a fitted-parameter-then-predicted result. The more serious gap is that BESSs are specified only in MVA and modeled via a 'stiff DC voltage source representing a battery energy storage system,' with no MWh rating, state-of-charge dynamics, or PV generation profile; this makes the energy-duration aspect of 'reliably support data center operations' over 24–36 months unsupported, but that is an omitted energy-budget check rather than an equivalence-by-construction, so it does not itself constitute circularity. The only clear self-citation-loaded element is Case 7, where the GFM-control comparison and the adaptive-VSG recommendation are taken from the authors' own prior publications ([59], [60], [64]) and then presented as this manuscript's assessment. Those prior works are published and externally checkable, so this is not a uniqueness-theorem lock-in, but as presented it is a self-citation chain supporting a stated contribution. Accordingly, the score is 4: some self-citation with the central phased-framework/islanding claim still independently supported by the EMT results of Cases 1–6.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper's central design choice — natural-gas turbines plus GFM BESS rated 200-350 MVA serving a 300 MW islanded plant — rests on five unvalidated premises: unvalidated load-profile scaling, unstated battery energy capacity, standard-but-unverified EMT component models, transfer of the authors' earlier GFM-control comparisons into the AI data center restoration context, and unsourced industry lead-time values. No new physical entities are invented; the framework recombines existing components.

free parameters (5)
  • GFM droop coefficients m_p, n_q = m_p = 3.14e-6 rad/W·s; n_q = 55.2e-6 V/Var
    Table VIII. Chosen controller gains with no tuning or sensitivity method stated; standard ranges for 1 MW units but they set the frequency/voltage response that the paper's central conclusions rest on.
  • BESS power ratings per case = 200 MVA (Case 3), 350 MVA (Case 4), 300 MVA (Case 5), 300 MVA (Case 6)
    Table VII. Hand-picked so that recommended cases pass; no sizing calculation tied to the load profile's statistics; energy (MWh) never given.
  • Gas turbine fleet sizes = 10, 9, 4, 2 units of 40 MVA (Cases 1, 2, 3, 6)
    Table VII. Chosen per scenario. Case 1's claim that 8×40 MVA 'arithmetically meet' 300 MW ignores MVA vs MW and power factor (~288 MW at PF 0.9).
  • Load scaling multiplier from single-GB200 profile to 300 MW plant = unstated
    §V. The 300 MW profile is a proportional scaling of Fig. 4 with no aggregation model, no diversification, and no validation against hyperscale measurements.
  • Acceptance thresholds ('tight' voltage/frequency bounds) = 0.95-1.05 p.u.; 59.9-60.25 Hz
    §V Cases 4-5. The thresholds the paper uses to declare success are chosen by the authors without reference to a defined standard for AI training loads.
axioms (5)
  • domain assumption A single-GB200 training profile (Fig. 4, Mistral 7B) scaled linearly represents a 300 MW hyperscale AI training facility.
    §V intro: 'emulated and scaled to represent a data center plant load of 300 MW'. Unvalidated aggregation of GPU diversity, checkpointing, and queuing.
  • domain assumption The islanded BESS-PV systems have sufficient stored energy for the simulated horizon and for the stated 24-36-month islanded phases.
    Tables VII-VIII and §III-A give MVA ratings only; no MWh, no PV generation model, no solar-resource assumption. The binding constraint for islanded weeks-long operation is energy, not power.
  • standard math Standard EMT models (synchronous machine subtransient/transient reactances, SRF-PLL, droop APC/RPC) correctly represent the hardware.
    §IV and Table VIII. Standard practice, but unverified here; no model validation against measurements or hardware-in-the-loop tests (the paper itself recommends HIL in §VI-C).
  • domain assumption GFM-control comparative performance from [59], [60], [64] transfers unchanged to the AI data center restoration scenario.
    §V-G, Fig. 13 caption cites [59]; Table IX is a qualitative restatement of the same group's comparison. No AI-load-specific restoration simulation is run in this paper.
  • domain assumption Procurement lead times and phase durations in Table II reflect current market conditions.
    Table II and §III-A. Values are stated as 'authors' compilation based on industry experience' without a cited source.

pith-pipeline@v1.3.0-alltime-deepseek · 24291 in / 17607 out tokens · 177798 ms · 2026-08-01T18:06:45.330362+00:00 · methodology

0 comments
read the original abstract

As hyperscale and colocation AI data centers continue to expand, the electric grid is increasingly required to support large, concentrated loads, with individual facilities ranging from 500 MW to 2 GW. Current projections estimate that approximately 50 GW of AI data center capacity will require grid connectivity in the United States by 2030. While prior research has extensively examined the environmental and operational impacts of AI data centers, as well as their potential role as grid-interactive assets, limited attention has been given to the challenges associated with their scalable deployment through engineering, procurement, and construction (EPC) processes. This manuscript addresses this gap by proposing a phased development framework for AI data center expansion. The approach is designed to enable developers to meet aggressive time-to-market objectives while navigating multi-year constraints associated with interconnection approvals and lead times associated with the procurement of component equipment. A modular construction architecture is presented, along with a detailed analysis of integrated energy systems and the role of hybrid on-site generation in supporting incremental capacity growth. Electromagnetic transient simulations (EMT) are used to evaluate system performance, demonstrating that a combination of on-site natural gas generation and grid-forming energy storage can reliably support data center operations during early and intermediate deployment phases. The study further examines the transition to full grid interconnection, including the capability of the data center to operate in islanded mode during grid disturbances. Finally, the manuscript compares grid-forming control strategies for system reconnection and restoration under varying conditions.

Figures

Figures reproduced from arXiv: 2607.17391 by Mohammad Ashraf Hossain Sadi, Nabil Mohammed, Soham Ghosh.

Figure 1
Figure 1. Figure 1: Forecasted capacity of US AI data centers under various project scenarios. Source: epoch.ai. [9] [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The interconnection process defined under FERC Order 2023 incorporates escalating study deposits and withdrawal [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Proposed phased development framework for AI data center construction with initial grid interconnection constraints. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: GPU power consumption for different stages of AI [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Year-over-year trend for containerized LFP BESS [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Hybrid energy system combinations for on-site AI data center deployment. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Electric circuit and control structures of the investigated IBRs [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Illustration of case 1 - Operation of the data center plant with on-site gas turbines generation. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Illustration of case 3 - Operation of the data center plant with GFM BESS-PV system and on-site gas turbines [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Illustration of case 4 - Operation of the data center plant with GFM BESS-PV systems. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Illustration of case 5 - Operation of the data center plant with GFM and GFL BESS-PV systems. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Illustration of case 6 - Operation of the data center plant in grid connected mode and transition into islanded on-site [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Performance of the GFMIs in a strong grid with SCR = 8 and [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗

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

64 extracted references · 6 linked inside Pith

  1. [1]

    Generative arti- ficial intelligence: a systematic review and applications,

    S. S. Sengar, A. B. Hasan, S. Kumar, and F. Carroll, “Generative arti- ficial intelligence: a systematic review and applications,”Multimedia Tools and Applications, vol. 84, no. 21, pp. 23 661–23 700, 2025

    2025

  2. [2]

    The potential of generative artificial intelligence across disciplines: Perspectives and future directions,

    K.-B. Ooi, G. W.-H. Tan, M. Al-Emran, M. A. Al-Sharafi, A. Capatina, A. Chakraborty, Y . K. Dwivedi, T.-L. Huang, A. K. Kar, V .-H. Leeet al., “The potential of generative artificial intelligence across disciplines: Perspectives and future directions,”Journal of Computer Information Systems, vol. 65, no. 1, pp. 76–107, 2025

    2025

  3. [3]

    A survey of large language models,

    W. X. Zhao, K. Zhou, J. Li, T. Tang, Z. Dong, Y . Hou, B. Zhang, Y . Min, J. Zhang, P. Liuet al., “A survey of large language models,” Frontiers of Computer Science, vol. 20, no. 12, p. 2012627, 2026

    2026

  4. [4]

    Large multimodal agents: A survey,

    J. Xie, Z. Chen, R. Zhang, X. Wan, and G. Li, “Large multimodal agents: A survey,”arXiv preprint arXiv:2402.15116, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Current and future applications of artificial intelligence in power systems: A critical appraisal,

    R. J. Bessa, S. Chatzivasileiadis, N. Zhang, C. Kang, and N. Hatziar- gyriou, “Current and future applications of artificial intelligence in power systems: A critical appraisal,”Journal of Modern Power Sys- tems and Clean Energy, vol. 14, no. 1, pp. 23–36, 2026

    2026

  6. [6]

    What is making artificial intelligence so successful today: a new generation of chips, algorithms, and toolboxes,

    A. Vlachos, A. Poulopoulou, C. Giannoula, G. Goumas, and N. Koziris, “What is making artificial intelligence so successful today: a new generation of chips, algorithms, and toolboxes,”Journal of Modern Power Systems and Clean Energy, vol. 14, no. 1, pp. 7–22, 2026

    2026

  7. [7]

    Immersion cooling fluid comparison using a single-phase immersion/liquid technique for data centers,

    A. Ruichek, M. Hnayno, A. Chehade, F. Fardoun, T. Moussa, G. Poli- dori, and C. Maalouf, “Immersion cooling fluid comparison using a single-phase immersion/liquid technique for data centers,”Applied Thermal Engineering, p. 128355, 2025

    2025

  8. [8]

    Scaling intelligence: The exponential growth of ai’s power needs,

    J. You, D. Owen, D. Porter, and T. Wilson, “Scaling intelligence: The exponential growth of ai’s power needs,”EPRI, 2025

    2025

  9. [9]

    How much power will frontier ai training demand in 2030?

    J. You and D. Owen, “How much power will frontier ai training demand in 2030?” https://epoch.ai/publications/ power-demands-of-frontier-ai-training, 2025, [Accessed 05-14- 2026]

    2030

  10. [10]

    Ai, data centers and the coming us power demand surge,

    C. Davenport, B. Singer, N. Mehta, B. Lee, J. Mackay, A. Modak, B. Corbett, J. Miller, T. Hari, J. Ritchieet al., “Ai, data centers and the coming us power demand surge,”Goldman Sachs, vol. 26, 2024

    2024

  11. [11]

    Intelligent resource management: Ai methods for predictive workload forecasting in cloud data centers,

    S. Padur, “Intelligent resource management: Ai methods for predictive workload forecasting in cloud data centers,”J. Artif. Intell. Mach. Learn. & Data Sci, vol. 1, no. 1, pp. 2936–2941, 2022

    2022

  12. [12]

    Evolution of data center design to handle ai workloads,

    T. Gupta, “Evolution of data center design to handle ai workloads,” in2024 34th International Telecommunication Networks and Applica- tions Conference (ITNAC). IEEE, 2024, pp. 1–8

    2024

  13. [13]

    Ai- optimized hyperscale data centers: Meeting the rising demands of generative ai workloads,

    T. Sankar, R. B. Venkata Ramana, and A. Balamuralikrishnan, “Ai- optimized hyperscale data centers: Meeting the rising demands of generative ai workloads,”International Journal of Trend in Scientific Research and Development, vol. 7, no. 1, pp. 1504–1514, 2023

    2023

  14. [14]

    New trends in photonic switching and optical networking architectures for data centers and computing systems,

    S. B. Yoo, “New trends in photonic switching and optical networking architectures for data centers and computing systems,”Journal of Optical Communications and Networking, vol. 15, no. 8, pp. C288– C298, 2023

    2023

  15. [15]

    The rising energy footprint of data centers: A review of global trends, challenges, and opportunities,

    F. A. Nahid, M. M. Islam, M. Kamruzzaman, and A. Al Mansur, “The rising energy footprint of data centers: A review of global trends, challenges, and opportunities,” in2025 IEEE 7th International Conference on Sustainable Technologies For Industry 5.0 (STI). IEEE, 2025, pp. 1–6

    2025

  16. [16]

    Carbon footprint of ai data centers: a life cycle approach,

    A. d’Orgeval, E. Assoumou, V . Sessa, I. Colak, S. Sheehan, and Q. Avenas, “Carbon footprint of ai data centers: a life cycle approach,” inProceedings of the 16th International Conference on Applied Energy, 2024, pp. 112–123

    2024

  17. [17]

    The carbon and water footprints of data centers and what this could mean for artificial intelligence,

    A. de Vries-Gao, “The carbon and water footprints of data centers and what this could mean for artificial intelligence,”Patterns, vol. 7, no. 1, 2026

    2026

  18. [18]

    Gone with the clouds: Estimating the electricity and water footprint of digital data services in europe,

    J. Farfan and A. Lohrmann, “Gone with the clouds: Estimating the electricity and water footprint of digital data services in europe,” Energy Conversion and Management, vol. 290, p. 117225, 2023

    2023

  19. [19]

    Ai data centres as grid-interactive assets,

    P. Colangelo, A. K. Coskun, J. Megrue, C. Roberts, S. Sengupta, V . Sivaram, E. Tiao, A. Vijaykar, C. Williams, D. C. Wilsonet al., “Ai data centres as grid-interactive assets,”Nature Energy, vol. 11, no. 2, pp. 254–261, 2026

    2026

  20. [20]

    Exploding ai power use: an opportunity to rethink grid planning and management,

    L. Lin, R. Wijayawardana, V . Rao, H. Nguyen, E. W. GNIBGA, and A. A. Chien, “Exploding ai power use: an opportunity to rethink grid planning and management,” inProceedings of the 15th ACM International Conference on Future and Sustainable Energy Systems, 2024, pp. 434–441

    2024

  21. [21]

    Mitigating power grid impact from proactive data center workload shifts: A coordinated scheduling strategy integrating synergistic traffic-data-power networks,

    Y . Zhang, B. Zou, X. Jin, Y . Luo, M. Song, Y . Ye, Q. Hu, Q. Chen, and A. C. Zambroni, “Mitigating power grid impact from proactive data center workload shifts: A coordinated scheduling strategy integrating synergistic traffic-data-power networks,”Applied Energy, vol. 377, p. 124697, 2025

    2025

  22. [22]

    Mitigating curtailment and carbon emissions through load migration between data centers,

    J. Zheng, A. A. Chien, and S. Suh, “Mitigating curtailment and carbon emissions through load migration between data centers,”Joule, vol. 4, no. 10, pp. 2208–2222, 2020

    2020

  23. [23]

    The unseen ai dis- ruptions for power grids: Llm-induced transients,

    Y . Li, M. Mughees, Y . Chen, and Y . R. Li, “The unseen ai dis- ruptions for power grids: Llm-induced transients,”arXiv preprint arXiv:2409.11416, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Ai-driven data center energy profile, power quality, sustainable sitting, and energy management: A comprehensive survey,

    R. R. Ahrabi, A. Mousavi, E. Mohammadi, R. Wu, and A. K. Chen, “Ai-driven data center energy profile, power quality, sustainable sitting, and energy management: A comprehensive survey,” in2025 IEEE Conference on Technologies for Sustainability (SusTech). IEEE, 2025, pp. 1–8

    2025

  25. [25]

    Data center control against sub-synchronous resonance: A data-driven approach,

    G. Ruan, M. D. Ilic, and L. Xie, “Data center control against sub-synchronous resonance: A data-driven approach,”arXiv preprint arXiv:2511.14141, 2025

    arXiv 2025

  26. [26]

    Using grid-forming energy storage systems to provide dynamic active power support for hyperscale data center,

    B. A. Ross, X. Lyu, U. C. Nwaneto, S. M. Mohiuddin, and A. B. Nas- sif, “Using grid-forming energy storage systems to provide dynamic active power support for hyperscale data center,”IEEE Access, 2026

    2026

  27. [27]

    Ai data center load forecasting: A liquid neural network approach,

    M. Mughees, Y . Li, Z. Lin, Y . Chen, and Y . R. Li, “Ai data center load forecasting: A liquid neural network approach,”IEEE Transactions on Smart Grid, 2026

    2026

  28. [28]

    Learning from imbal- ance: Cross-server power prediction in large data centers via domain adaptation regression,

    R. Mo, W. Lin, G. Liu, H. Liu, and L. He, “Learning from imbal- ance: Cross-server power prediction in large data centers via domain adaptation regression,”Expert Systems with Applications, vol. 287, p. 127845, 2025

    2025

  29. [29]

    Scalable data centers–power generation and delivery challenges and solutions,

    S. Ghosh and J. Damodara, “Scalable data centers–power generation and delivery challenges and solutions,”Energy Reports, vol. 15, p. 109178, 2026

    2026

  30. [30]

    From barrier to bridge: The case for ai data center/power grid co-design,

    N. Bashir, R. Sherwood, L. Xie, and M. Yu, “From barrier to bridge: The case for ai data center/power grid co-design,”arXiv preprint arXiv:2605.03090, 2026

    Pith/arXiv arXiv 2026

  31. [31]

    Practical guidance and considerations for large load interconnections,

    R. Quint, K. Thomas, J. Zhao, A. Isaacs, and C. Baker, “Practical guidance and considerations for large load interconnections,”Elevate Energy Consulting, Technical Report, May, 2025

    2025

  32. [32]

    Grid connection barriers to renewable energy deployment in the united states,

    W. Gorman, J. M. Kemp, J. Rand, J. Seel, R. Wiser, N. Manderlink, F. Kahrl, K. Porter, and W. Cotton, “Grid connection barriers to renewable energy deployment in the united states,”Joule, vol. 9, no. 2, 2025

    2025

  33. [33]

    Data on AI Models,

    Epoch AI, “Data on AI Models,” 5 2026, accessed: 17 May 2026. [Online]. Available: https://epoch.ai/data/ai-models

    2026

  34. [34]

    Ai load dynamics–a power electronics perspective,

    Y . Li and Y . Li, “Ai load dynamics–a power electronics perspective,” arXiv preprint arXiv:2502.01647, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Short-term load forecasting for ai-data center,

    M. Mughees, Y . Li, Y . Chen, and Y . R. Li, “Short-term load forecasting for ai-data center,” in2025 IEEE Power & Energy Society General Meeting (PESGM). IEEE, 2025, pp. 1–5

    2025

  36. [36]

    Power stabilization for ai training datacenters,

    E. Choukse, B. Warrier, S. Heath, L. Belmont, A. Zhao, H. A. Khan, B. Harry, M. Kappel, R. J. Hewett, K. Dattaet al., “Power stabilization for ai training datacenters,”arXiv preprint arXiv:2508.14318, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Sustainable supercomputing for ai: Gpu power capping at hpc scale,

    D. Zhao, S. Samsi, J. McDonald, B. Li, D. Bestor, M. Jones, D. Tiwari, and V . Gadepally, “Sustainable supercomputing for ai: Gpu power capping at hpc scale,” inProceedings of the 2023 ACM symposium on cloud computing, 2023, pp. 588–596

    2023

  38. [38]

    Ai data centers: A review of power quality challenges and mitigation strategies,

    L. Maheshwari, H. Dai, B. Sarlioglu, R. Chakraborty, Y . Xia, and M. Harper, “Ai data centers: A review of power quality challenges and mitigation strategies,” in2026 IEEE Power and Energy Conference at Illinois (PECI). IEEE, 2026, pp. 1–6

    2026

  39. [39]

    Fine-tuning ai models with limited resources,

    K. Singhapoo, A. Inthanil, and A. Pillai, “Fine-tuning ai models with limited resources,” in2025 11th International Conference on Engineering, Applied Sciences, and Technology (ICEAST). IEEE, 2025, pp. 148–151

    2025

  40. [40]

    Assessing the carbon footprint of language models: Towards sustainability in ai,

    F. Jeanquartier, C. Jean-Quartier, P. Rieder, V . Misirli ´c, C. Pasero, R. Hohensinner, H. M ¨uller, and A. Holzinger, “Assessing the carbon footprint of language models: Towards sustainability in ai,”Resources, Conservation and Recycling, vol. 226, p. 108670, 2026

    2026

  41. [41]

    How green are neural language models? analyzing energy consumption in text sum- marization fine-tuning,

    T. Rehman, D. K. Sanyal, and S. Chattopadhyay, “How green are neural language models? analyzing energy consumption in text sum- marization fine-tuning,” inInternational Conference on Power Engi- neering and Intelligent Systems (PEIS). Springer, 2025, pp. 649–661

    2025

  42. [42]

    Electricity demand and grid impacts of ai data centers: Challenges and prospects,

    X. Chen, X. Wang, A. Colacelli, M. Lee, and L. Xie, “Electricity demand and grid impacts of ai data centers: Challenges and prospects,” arXiv preprint arXiv:2509.07218, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Power hungry processing: Watts driving the cost of ai deployment?

    S. Luccioni, Y . Jernite, and E. Strubell, “Power hungry processing: Watts driving the cost of ai deployment?” inProceedings of the 2024 ACM conference on fairness, accountability, and transparency, 2024, pp. 85–99

    2024

  44. [44]

    Battery technologies for grid-scale energy storage,

    T. Jiang, D. Shen, Z. Zhang, H. Liu, G. Zhao, Y . Wang, S. Tan, R. Luo, and W. Chen, “Battery technologies for grid-scale energy storage,” Nature Reviews Clean Technology, vol. 1, no. 7, pp. 474–492, 2025

    2025

  45. [45]

    Recent advancements in membrane- free redox flow batteries,

    X. Wang, R. K. Gautamet al., “Recent advancements in membrane- free redox flow batteries,”Chemical Society Reviews, vol. 54, no. 12, pp. 5895–5911, 2025

    2025

  46. [46]

    Application of superconducting magnetic energy storage in electrical power and energy systems: a review,

    V . S. Vulusala G and S. Madichetty, “Application of superconducting magnetic energy storage in electrical power and energy systems: a review,”International Journal of Energy Research, vol. 42, no. 2, pp. 358–368, 2018

    2018

  47. [47]

    Superconduct- ing magnetic energy storage systems: Prospects and challenges for renewable energy applications,

    B. B. Adetokun, O. Oghorada, and S. J. Abubakar, “Superconduct- ing magnetic energy storage systems: Prospects and challenges for renewable energy applications,”Journal of Energy Storage, vol. 55, p. 105663, 2022

    2022

  48. [48]

    An overview of smes applications in power and energy systems,

    M. H. Ali, B. Wu, and R. A. Dougal, “An overview of smes applications in power and energy systems,”IEEE transactions on sustainable energy, vol. 1, no. 1, pp. 38–47, 2010

    2010

  49. [49]

    Smes for power utility applications: A review of technical and cost considerations,

    V . Karasik, K. Dixon, C. Weber, B. Batchelder, G. Campbell, and P. Ribeiro, “Smes for power utility applications: A review of technical and cost considerations,”IEEE transactions on applied superconduc- tivity, vol. 9, no. 2, pp. 541–546, 1999

    1999

  50. [50]

    Modular multilevel converters: Recent achievements and challenges,

    M. A. Perez, S. Ceballos, G. Konstantinou, J. Pou, and R. P. Aguilera, “Modular multilevel converters: Recent achievements and challenges,” IEEE Open Journal of the Industrial Electronics Society, vol. 2, pp. 224–239, 2021

    2021

  51. [51]

    Energy storage systems for ai data centers: A review of technologies, characteristics, and applicability,

    S. Rahman and T. A. Khan, “Energy storage systems for ai data centers: A review of technologies, characteristics, and applicability,” Energies, vol. 19, no. 3, p. 634, 2026

    2026

  52. [52]

    Review of energy storage capacitor technology,

    W. Liu, X. Sun, X. Yan, Y . Gao, X. Zhang, K. Wang, and Y . Ma, “Review of energy storage capacitor technology,”Batteries, vol. 10, no. 8, p. 271, 2024

    2024

  53. [53]

    Battery energy storage systems: A review of energy management systems and health metrics,

    S. Nazaralizadeh, P. Banerjee, A. K. Srivastava, and P. Famouri, “Battery energy storage systems: A review of energy management systems and health metrics,”Energies, vol. 17, no. 5, p. 1250, 2024

    2024

  54. [54]

    A review of battery energy storage systems and advanced battery management system for different applications: Challenges and recommendations,

    S. Nyamathulla and C. Dhanamjayulu, “A review of battery energy storage systems and advanced battery management system for different applications: Challenges and recommendations,”Journal of Energy Storage, vol. 86, p. 111179, 2024

    2024

  55. [55]

    A review of lithium-ion battery state of charge estimation and management system in electric vehicle applications: Challenges and recommenda- tions,

    M. A. Hannan, M. H. Lipu, A. Hussain, and A. Mohamed, “A review of lithium-ion battery state of charge estimation and management system in electric vehicle applications: Challenges and recommenda- tions,”Renewable and Sustainable Energy Reviews, vol. 78, pp. 834– 854, 2017

    2017

  56. [56]

    Comprehensive review of energy storage systems technologies, objec- tives, challenges, and future trends,

    D. A. Elalfy, E. Gouda, M. F. Kotb, V . Bure ˇs, and B. E. Sedhom, “Comprehensive review of energy storage systems technologies, objec- tives, challenges, and future trends,”Energy Strategy Reviews, vol. 54, p. 101482, 2024

    2024

  57. [57]

    Lebowitz, S

    D. Lebowitz, S. Sundaram, and S. Daly,The BESS book - A cell to grid guide to utility-scale battery energy storage. ISBN-13: 979- 8218447984, 2024

    2024

  58. [58]

    Frequency stability of synchronous machines and grid-forming power converters,

    A. Tayyebi, D. Groß, A. Anta, F. Kupzog, and F. D ¨orfler, “Frequency stability of synchronous machines and grid-forming power converters,” IEEE Journal of Emerging and Selected Topics in Power Electronics, vol. 8, no. 2, pp. 1004–1018, 2020

    2020

  59. [59]

    Grid-forming inverters: A comparative study of different control strategies in frequency and time domains,

    N. Mohammed, H. Udawatte, W. Zhou, D. J. Hill, and B. Bahrani, “Grid-forming inverters: A comparative study of different control strategies in frequency and time domains,”IEEE Open Journal of the Industrial Electronics Society, vol. 5, pp. 185–214, 2024

    2024

  60. [60]

    Comparative frequency-and time-domain analysis of four grid-forming inverters: Droop, vsm, matching, and dvoc,

    N. Mohammed, S. Ahmed, and C. Konstantinou, “Comparative frequency-and time-domain analysis of four grid-forming inverters: Droop, vsm, matching, and dvoc,”International Journal of Electrical Power & Energy Systems, vol. 172, p. 111302, 2025

    2025

  61. [61]

    Hierarchical control of droop-controlled ac and dc micro- grids—a general approach toward standardization,

    J. M. Guerrero, J. C. Vasquez, J. Matas, L. G. De Vicu ˜na, and M. Castilla, “Hierarchical control of droop-controlled ac and dc micro- grids—a general approach toward standardization,”IEEE Transactions on industrial electronics, vol. 58, no. 1, pp. 158–172, 2010

    2010

  62. [62]

    Accurate reactive power sharing strategy for droop-based islanded ac microgrids,

    N. Mohammed, A. Lashab, M. Ciobotaru, and J. M. Guerrero, “Accurate reactive power sharing strategy for droop-based islanded ac microgrids,”IEEE Transactions on Industrial Electronics, vol. 70, no. 3, pp. 2696–2707, 2022

    2022

  63. [63]

    Incident review: Considering simultaneous voltage-sensitive load reductions,

    North American Electric Reliability Corporation (NERC), “Incident review: Considering simultaneous voltage-sensitive load reductions,” https://www.nerc.com/globalassets/our-work/reports/event-reports/ incident review large load loss.pdf, 2025, [Accessed 05-14-2026]

    2025

  64. [64]

    Online grid impedance estimation-based adaptive control of virtual synchronous generators considering strong and weak grid conditions,

    N. Mohammed, M. H. Ravanji, W. Zhou, and B. Bahrani, “Online grid impedance estimation-based adaptive control of virtual synchronous generators considering strong and weak grid conditions,”IEEE Trans- actions on Sustainable Energy, vol. 14, no. 1, pp. 673–687, 2022

    2022