{"id":"94757bef-8c19-4eff-bc42-185c944caac9","arxiv_id":"2505.16575","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A dynamic model of data center loads, including UPS trip/reconnect logic, cooling motors, and pulsing AI workloads, is demonstrated on a model of the Irish power grid.","lead":"This paper proposes a dynamic load model for data centers that includes uninterruptible power supply behavior, cooling motors, and pulsing AI workloads, and tests it on a model of the Irish grid. It aims to help grid operators predict sudden data-center disconnections and reconnections, including a repeated on-off pattern called flapping.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim that the model 'reproduces real-world observations' is unsupported: no comparison to measured events, and the AI-load square-wave assumption is not validated.","rationale":"The paper has a clear, implementable model structure with explicit equations and a realistic test system (the all-island Irish model in Dome). The UPS modes, reconnection logic, cooling motor, and filters are a reasonable engineering synthesis. My concern is not internal inconsistency: the equations are coherent as far as they go. The problem is evidential. The abstract and Section I-C promise a model 'adequate to reproduce real-world observations,' but the manuscript never performs that reproduction. The real-world data referenced are the DC location map and a motivating measured frequency trace in Fig. 1; the case studies are synthetic scenarios with assumed parameters. Without a comparison to a measured event, the central claim is not established. The AI-load assumption in Eq. (5) is the sharpest place where the model could be wrong: lockstep square waves across all GPUs/TPUs is a strong claim, and the low-pass filters (6)-(7) do not model aggregate desynchronization. But this is a component-level concern; the validation gap is the load-bearing one because even if the component is right, the whole model's fit to reality is unshown, and if the whole model is validated against a real event, the component concern can be calibrated. I therefore partially agree with the reader, whose weakest_assumption was the lockstep square wave; I think the no-validation issue is more decisive. The verdict should remain CONDITIONAL: the model is a plausible candidate, but acceptance should require the validation check described above.","tokens_in":13135,"tokens_out":5666,"duration_ms":48407,"concrete_test":"Replay the measured Kellystown-Woodland fault shown in Fig. 1: connect the proposed DC model (using the reported pre-fault DC demand and Section IV parameter values) at the affected Irish-system bus, apply the same fault clearing sequence, and compare simulated frequency and RoCoF against the PMU trace in Fig. 1. Set an acceptance threshold before running, e.g., simulated frequency nadir and maximum RoCoF within 10% of the measured values; if the model misses this, the central claim of reproducing real-world observations is not supported and the paper should be revised to present the model as a hypothesis rather than a validated model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section I-C promise a model that 'properly captures the behaviour of DCs' and is 'adequate to reproduce real-world observations,' but the paper never tests that claim. The only real-world elements are the DC location map and the motivating frequency trace of Fig. 1; the case studies in Section IV use a hypothetical 300 MW DC with assumed parameters. No simulated response is compared with a measured event, no parameter identification is performed, and no sensitivity analysis shows that the qualitative results are robust. In addition, the AI-load component rests on Eq. (5), a perfectly periodic square wave in u_GPU, justified by the assertion that all GPUs/TPUs activate and deactivate in lockstep. The low-pass filters in Eqs. (6)-(7) smooth individual server transients but do not represent inter-server desynchronization; if real AI training loops are staggered across servers, the aggregate load is a smoother, lower-amplitude signal than Eq. (5), and the frequency/voltage swings in Fig. 13 would be smaller. Either issue alone means the central claim is not established; together they leave the model as an interesting hypothesis rather than a validated load model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a dynamic load model of data centers for transient stability analysis. The model combines a UPS with fault-ride-through logic (voltage/frequency disconnection thresholds and delayed reconnection), an induction-motor cooling load, an aggregated ZIP load for miscellaneous equipment, and server loads split into CPU and GPU components, with the GPU load represented as a periodic square-wave pulse for AI workloads. The model is demonstrated on a dynamic model of the all-island Irish transmission system, using the real geographic distribution of Irish data centers, through scenarios covering constant load under a fault, flapping and segmented reconnection, batched and AI loads, periodic transients, and detailed reconnection with voltage-angle and flux-dynamics effects. The central claim is that the model 'properly captures the behaviour of DCs' and is 'adequate to reproduce real-world observations and anticipate potential instabilities, e.g., flapping' (Abstract; Section I-C).","tokens_in":13463,"tokens_out":5546,"duration_ms":47953,"significance":"If the validation claim were fully supported, the contribution would be timely and practically relevant: transmission system operators currently use generic PQ or voltage-dependent load models that cannot represent UPS disconnection/reconnection, cooling-motor transients, or AI-driven power pulses. The paper is transparent in construction: the model is assembled from standard component equations (ZIP load, induction motor, first-order filters, and exogenous stochastic processes), and there is no circularity, since the parameters are chosen rather than identified from a target output. The case studies on a realistic Irish system model illustrate plausible mechanisms and reproduce the qualitative phenomenon of flapping. The main risk is that the central claim is broader than the evidence: no simulated response is compared with a measured event, and the AI square-wave assumption is not empirically supported. As it stands, the manuscript is a credible model proposal rather than a validated load model, and the significance of the paper would be materially increased by adding field-data comparison or a clearly scoped validation study.","major_comments":[{"comment":"The abstract and Section I-C claim that the model 'properly captures the behaviour of DCs' and is 'adequate to reproduce real-world observations and anticipate potential instabilities.' This is the central claim, yet Section IV never tests it against the real event described in the Introduction. The 204 MW / 0.12 Hz/s event of Fig. 1 is used only as motivation; all case studies use a hypothetical 300 MW DC with assumed parameters, and no simulated frequency or power waveform is compared with a measured trace. The authors should either add a direct comparison with the cited event (or another field measurement) and a parameter-identification or sensitivity study, or alternatively re-scope the Abstract and Section I-C to present the contribution as a model proposal rather than a validated model. As written, the validation claim is not established.","section":"Abstract; Section I-C; Section IV"},{"comment":"The AI-load component rests on Eq. (5), a perfectly periodic square wave in u_GPU, justified in Section II-C by the assertion that 'Training of neural networks requires that all servers work in lockstep, therefore all GPUs and TPUs activate and deactivate at the same time.' Section III-A immediately qualifies this with 'each server's load might be unequal and/or its response not properly synchronized,' but the low-pass filters in Eqs. (6)-(7) only smooth within-server transients; they do not represent inter-server desynchronization or the aggregation of many servers with staggered training phases. If real AI workloads are even partially desynchronized across servers, the aggregate demand would be a smoother, lower-amplitude signal than Eq. (5), and the frequency and voltage swings in Fig. 13 would be smaller. The authors should either support the lockstep assumption with measurements of aggregate AI-DC demand or extend the model with an explicit aggregation/dephasing mechanism and show the sensitivity of the Section IV-C results to this assumption.","section":"Section II-C, Eq. (5); Section III-A, Eqs. (6)-(7); Section IV-C"},{"comment":"The flapping scenario is obtained by choosing specific values that are not justified or varied: the DC is increased to roughly 420 MW by 'double cooling load,' the reconnection delay is set to 10 s, and the disturbance-counting scheme is not active. The text itself acknowledges that flapping risk depends on DC size and grid loading, but no sensitivity analysis is presented. Because the paper's claim to 'anticipate potential instabilities' depends on this phenomenon, the robustness of the flapping prediction to parameter uncertainty should be characterized; otherwise the statement that the cycle 'can repeat indefinitely in the worst scenario' is not quantified.","section":"Section IV-B"}],"minor_comments":[{"comment":"The variables Delta_f and Delta_v in Eq. (16) are not defined. As written, the first inequality (Delta_f < f_min) is satisfied for a null frequency deviation, which cannot be the intended disconnection condition. Please clarify whether the intended logic is f < f_min, f > f_max, v < v_min, v > v_max, or an equivalent absolute-deviation formulation.","section":"Section III-F, Eq. (16)"},{"comment":"The text uses 'disconnection time' to mean the delay before reconnection (30 s and 10 s in the two scenarios), although the disconnection itself is assumed instantaneous. The term 'reconnection delay' would avoid confusion with the UPS trip time.","section":"Section IV-A and Section IV-B"},{"comment":"The statement that the back-up generator is 'implicitly embedded in the UPS' is a simplification that should be listed as a limitation. In particular, the model as written cannot represent the finite energy of the UPS/battery or the start-up delay and dynamics of the back-up generator, which may be relevant in extended FRT events.","section":"Section III-D"},{"comment":"The jump-diffusion process for u_CPU in Eq. (3) is unbounded, whereas u_CPU should remain in [0,1]. If the compound Poisson jumps can take arbitrary amplitudes, the model can produce out-of-range values unless a saturation or a bounded jump distribution is imposed. Please state the bounds used in the simulations.","section":"Section II-B, Eq. (3)"},{"comment":"Reference [19] (the aggregate DER model) lacks a publisher or report number; as cited, it is difficult to locate. Please provide a complete reference.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid model-construction paper with a clear and useful architecture, and the Irish-system case studies make the mechanisms concrete. The main gap is not circularity or internal inconsistency but the absence of a quantitative comparison with a measured event, together with an unvalidated lockstep assumption for AI loads. I would not reject the paper, because the model is reusable and the flapping mechanism is worth publishing, but the central validation claim in the Abstract and Section I-C must be either supported or explicitly narrowed. The self-citations [25] and [26] are peripheral and do not inflate the contribution. Fit to the journal is good; the revision should focus on validation and parameter sensitivity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a read if you work on load modeling for transient stability. It assembles the first aggregate grid-level DC model I know of that combines UPS disconnection/reconnection logic with reconnection delays, a cooling induction motor, ZIP auxiliary load, and a periodic GPU/TPU pulse for AI workloads, then runs it on a realistic all-island Irish system model with actual DC locations. The equations are standard and clearly presented; the flapping demo in Fig. 10 is a nice illustration of how delayed reconnection can go unstable, and the segmented reconnection fix in Fig. 11 is a practical point TSOs will care about.\n\nWhat it does not do is validate the central claim in the abstract and Section I-C that the model \"properly captures the behaviour of DCs\" and is \"adequate to reproduce real-world observations.\" No simulation is compared with the measured 204 MW event from [1] or with any other field data; all DC parameters are hand-picked, and there is no sensitivity analysis. That is a real gap, but it is an addressable one: the paper is best read as a credible model proposal, not a validated model. The same goes for the AI load. Equation (5) assumes all GPU/TPUs activate in lockstep, so the aggregate IT load is a sharp square wave. The authors filter each server's response but not inter-server desynchronization, so the predicted frequency swings in Fig. 13 may be an upper bound. If real AI training loops are staggered, the swings get smaller. That assumption should be justified or relaxed; it is the weakest single element of the model.\n\nI don't see circularity problems: no fitted quantity is used as a target, and the self-citations to [25],[26] are for standard stochastic integration tools and are peripheral. The literature review is honest about what DER_A lacks (no reconnection delay, no AI pulse), and the gap is real. Reproducibility is weaker than it should be: Dome is named, but input data and parameter files are not given, so a TSO could not rerun the cases as-is.\n\nBottom line: this deserves a serious referee. With validation against a measured event or a parameter-identification exercise, plus a stated justification for the square-wave assumption, it would be a solid contribution. I'd send it out.","headline":"A credible and clearly presented aggregate DC model for transient stability, but the abstract overclaims validation and the square-wave AI-load assumption is the main unguarded element.","tokens_in":13901,"tokens_out":1932,"would_cite":true,"duration_ms":17518,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dynamic load model with UPS switching, cooling-motor dynamics, and pulsing AI workload reproduces recorded data-center fault events and predicts a repeated disconnect–reconnect instability, called flapping, that occurs when reconnection…","keywords":["data center","load modeling","transient stability","fault ride-through","UPS","induction motor","AI workloads","flapping"],"falsifier":"Record the active power at the grid connection point of a real AI-training data center with sub-second sampling across several epochs: if the trace shows smooth ramps rather than sharp 0%-to-100% square-wave pulses, the pulse-train load model overpredicts frequency and voltage deviations.","tokens_in":12977,"feed_emoji":"⚡","tokens_out":8201,"duration_ms":63542,"temperature":0.7,"pith_summary":"The paper proposes a dynamic load model for data centers so that transmission system operators can simulate how large computing facilities behave during grid faults. The model joins three elements: an uninterruptible power supply with explicit disconnect-and-reconnect logic, a cooling system represented as an induction motor, and a pulsing load that mimics the periodic activation of AI accelerators. Tested on a detailed model of the all-island Irish transmission system using real data-center information, the model reproduces a real 204 MW demand drop and shows that fast reconnection after a fault can trigger flapping, a repeated cycle of disconnection and reconnection that pushes frequency below the protection threshold. The authors argue this model captures data-center dynamics that generic load models miss, making it a tool for anticipating such instabilities.","feed_headline":"New model reproduces data-center grid flapping","feed_subtitle":"UPS switching, cooling motors, and pulsing AI loads capture real 200-MW drops and reconnection risks.","key_machinery":"The central object is a mode-switching UPS model whose state (normal, emergency, or charging) is driven by voltage and frequency thresholds; in emergency mode the data center draws zero active and reactive power from the grid, and reconnection is gated by a delay or by disturbance-counting logic and by a phase-angle matching condition. The IT load is a sum of a CPU term driven by a compound-Poisson jump process and a GPU term modelled as a train pulse of active and idle periods, each smoothed by first-order low-pass filters representing power-supply transients. The cooling load is a dq-axis squirrel-cage induction motor with stator and rotor flux dynamics, and the remaining loads are a voltage-dependent ZIP model. These components are coupled through the UPS power-balance and stored-energy equations, which switch the grid-side power between normal, emergency, and charging expressions.","core_discovery":"The central claim is that a data center's grid interaction is controlled by three behaviours that standard load models omit: the UPS disconnects the whole facility when voltage or frequency leaves a defined band and reconnects only after a delay; the cooling load, about 30% of the total, draws power through an induction motor with flux dynamics; and AI training creates periodic square-wave demand as GPUs and TPUs switch between idle and active in lockstep. Put together, these features reproduce observed fault-ride-through events and expose a failure mode: when the reconnection delay is too short, the sudden load reconnection pushes frequency back below the trip threshold, disconnecting the data center again, and the cycle can repeat. The model also shows that a segmentation strategy where a large data center is split into several UPS units with staggered reconnection times eliminates flapping and creates a smooth ramp from zero to full load.","pith_inferences":["If GPU workloads are not perfectly synchronized across a facility, the aggregate demand is smoother than a square-wave train, so the frequency and voltage swings shown here are upper bounds; sub-second measurements of a real AI-DC power trace could test this.","The same UPS-plus-motor-plus-pulse architecture could be adapted to other large converter-coupled loads with fault-ride-through behaviour, such as electrolyzers and EV charging hubs.","The segmentation result points to a concrete engineering rule—require multiple UPSs with staggered reconnection delays for hyperscale data centers—which can be validated in hardware-in-the-loop before appearing in grid codes.","Because the model deliberately omits temperature and thermodynamic dynamics, it covers short-term transients only; assessing long-term frequency quality would require coupling it to a thermal model of the building."],"forward_implications":["Transmission system operators can simulate DC-rich grids and see realistic frequency, rate-of-change-of-frequency, and voltage excursions after faults, including the repeated disconnect–reconnect cycle called flapping.","Reconnection settings become a design lever: longer delays, or multiple UPS segments with staggered reconnection times, can eliminate flapping and smooth the return to full load.","AI data centers can cause frequency swings on the order of ±0.2 Hz during normal training cycles, and the swing size is reduced by increasing the power-supply time constant $T_{GPU}$.","The model gives a quantitative basis for setting grid-connection requirements, since it shows how DC size, grid loading, and reconnection logic interact to create instability risk."],"supporting_citations":[{"why":"supplies the real-world fault-ride-through event (204 MW drop, 0.12 Hz/s RoCoF) that motivates the model and defines its validation target.","marker":"[1]"},{"why":"documents the quick-burst power consumption pattern of large AI loads that the pulse-train GPU model is built to represent.","marker":"[22]"},{"why":"provides the CPU/GPU power consumption model with idle, full-load, and burst terms that the server load equations are based on.","marker":"[24]"},{"why":"gives the compound-Poisson jump process used to model irregular batched-task CPU load.","marker":"[25]"},{"why":"provides the stochastic differential equation framework used for the Ornstein-Uhlenbeck load noise term.","marker":"[26]"},{"why":"supplies the standard dq-axis induction-motor model with stator and rotor flux dynamics used for the cooling load.","marker":"[30]"},{"why":"documents the disturbance-counting and delayed-reconnection schemes that the UPS reconnection logic implements.","marker":"[31]"},{"why":"provides the dynamic simulation environment in which the all-island Irish system case studies are run.","marker":"[34]"},{"why":"introduces the flapping phenomenon of repeated load disconnection and reconnection that the paper reproduces.","marker":"[35]"}],"fun_headline_variants":["AI data centers flap the grid; model shows why","Staggered UPS reconnection stops data center flapping","Data center model captures AI-induced pulsing loads","Model explains data center grid flapping and its cure"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's AI-load predictions assume that all GPU and TPU accelerators in a facility turn on and off at the same time, so the combined load is a sharp square-wave pulse; if real workloads are staggered or software-smoothed, the predicted grid swings would be smaller.","fun_headline_variants_meta":{"raw":{"variants":["AI data centers flap the grid; model shows why","Staggered UPS reconnection stops data center flapping","Data center model captures AI-induced pulsing loads","Model explains data center grid flapping and its cure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001197,"raw_usage":{"total_tokens":4904,"prompt_tokens":882,"completion_tokens":4022,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":3959}},"tokens_in":498,"tokens_out":4022,"duration_ms":23845,"temperature":1.0,"reasoning_tokens":3959,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:58:08.446520+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the active power at the grid connection point of a real AI-training data center with sub-second sampling across several epochs: if the trace shows smooth ramps rather than sharp 0%-to-100% square-wave pulses, the pulse-train load model overpredicts frequency and voltage deviations.","supporting_citations":[{"cited_title":"Impact of Converter-based Demand on Frequency Quality in the Ireland and Northern Ireland Power Systems,","cited_arxiv_id":null,"evidence_quote":"supplies the real-world fault-ride-through event (204 MW drop, 0.12 Hz/s RoCoF) that motivates the model and defines its validation target."},{"cited_title":"An Assessment of Large Load Interconnection Risks in the Western Interconnection,","cited_arxiv_id":null,"evidence_quote":"documents the quick-burst power consumption pattern of large AI loads that the pulse-train GPU model is built to represent."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the CPU/GPU power consumption model with idle, full-load, and burst terms that the server load equations are based on."},{"cited_title":"Modeling solar irradiance for short-term dynamic analysis of power systems,","cited_arxiv_id":null,"evidence_quote":"gives the compound-Poisson jump process used to model irregular batched-task CPU load."},{"cited_title":"A systematic method to model power systems as stochastic differential algebraic equations,","cited_arxiv_id":null,"evidence_quote":"provides the stochastic differential equation framework used for the Ornstein-Uhlenbeck load noise term."},{"cited_title":"Kundur,Power System Stability and Control","cited_arxiv_id":null,"evidence_quote":"supplies the standard dq-axis induction-motor model with stator and rotor flux dynamics used for the cooling load."},{"cited_title":"Incident Review - Considering Simultaneous V oltage-Sensitive Load Reductions,","cited_arxiv_id":null,"evidence_quote":"documents the disturbance-counting and delayed-reconnection schemes that the UPS reconnection logic implements."},{"cited_title":"A Python-based software tool for power system analysis,","cited_arxiv_id":null,"evidence_quote":"provides the dynamic simulation environment in which the all-island Irish system case studies are run."},{"cited_title":"Micro-flexibility: Challenges for power system modeling and control,","cited_arxiv_id":null,"evidence_quote":"introduces the flapping phenomenon of repeated load disconnection and reconnection that the paper reproduces."}],"review_version":1}