Flexible Training Workloads in Large-Scale AI Data Centers for Transient-Stability Support in Transmission-Constrained Power Systems

arXiv:2608.30901 · eess.SY, cs.SY · Submitted 2026-08-31 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Flexible Training Workloads in Large-Scale AI Data Centers for Transient-Stability Support in Transmission-Constrained Power Systems".

Dev: The rapid expansion of large-scale artificial intelligence (AI) data centers is adding substantial, concentrated, and rapidly varying loads to transmission-constrained power systems.

Rosa: First, who's behind it and why it matters.

Summary of the Paper's Core Proposal and Immediate Implications: Rosa: Moving on to what the paper actually proposes in detail, it centers on Training-Induced Load Surge, or TILS, which is a fast demand-side strategy that kicks in after a fault clears by initiating or resuming flexible AI training workloads to increase active power demand at electrically effective locations.

Dev: That means we are looking at how the active power response of these AI workloads behaves during that critical post-fault window, and the paper shows this response is measured in terms of load (MW) and time (s), specifically showing a load response to training workload initiation four.

Taro: I'm interested in the quantitative data they present; they show the AI data center load response to a fifty MW AI data center in the United States, which gives us some concrete numbers to work with.

Rosa: They provide specific figures showing the load response, and this is important because it shows that flexible computing workloads can respond on timescales relevant to post-fault transient stability. The paper notes that these stabilizing mechanisms are dependent on response timing, magnitude, and electrical location.

Dev: That dependence on timing and location is exactly what worries me from a latency perspective; if the activation delay is too long or the siting isn't right, all that measured response could be lost because the instability has already progressed.

Taro: So, to summarize their findings, they’ve quantified how much power increase can be achieved by TILS and how that effectiveness changes based on when it happens and where it happens electrically.

Rosa: That's right; they quantify the effect of response magnitude, activation delay, and electrical siting across three systems: SMIB, IEEE thirty-nine-bus, and a large-scale Korean power system.

Dev: And those three system evaluations show that TILS can actually increase the transient-stability-constrained generation limit in all of them when conditions are right. That’s the core result we need to focus on for our loop rate analysis.

Discussing Suggested Improvements and Deeper Implications: Rosa: Now, let's look at what they suggest as improvements; they emphasize that TILS should be regarded as a complement to existing stability-enhancing measures rather than a replacement for them.

Dev: That makes sense from an engineering standpoint; we’re not trying to replace physical infrastructure or established controls with something completely new if it doesn't have the right reliability profile. The authors stress that deployment requires sufficient electrical headroom, flexible workloads, and reliable grid-triggered activation before a contingency occurs.

Taro: I think the implication here is that this isn't a magic fix; it depends entirely on having those specific conditions met before the event happens, which grounds the proposal in practical reality.

Rosa: Exactly; they also point out that deployment needs to be optimized by ensuring workloads are activated at electrically effective buses near critical generators with minimal activation delay, aiming for sub-second response times.

Dev: Sub-second is a tight target for us to hit, Rosa; we're dealing with component timescales that are estimated based on prior studies suggesting an aggregate response time around zero point one six seconds after fault clearing in some scenarios.

Taro: If we can meet those timing requirements, then the system transitions from being just theoretical and becomes something that could potentially be deployed alongside other methods.

Rosa: That’s the exciting part; it suggests that AI data centers could become a complementary resource when they have the right operational flexibility and grid triggers are reliable.

Conclusion: Dev: So, to wrap up on this discussion of "Flexible Training Workloads in Large-Scale AI Data Centers for Transient-Stability Support in Transmission-Constrained Power Systems," we’ve seen that TILS can potentially boost generation limits across different systems under the right conditions.

Rosa: It really boils down to using AI workloads as a fast demand response resource when we have sufficient electrical headroom and reliable grid triggers available to manage transient stability issues.

Taro: I think the implication is that the future of distributed energy management might involve coordinating flexible computing resources with power system operations in ways we haven't fully explored yet.

Dev: Agreed; it certainly adds a new dimension to our toolkit, but we still need rigorous validation before we can integrate anything into live control systems.

Rosa: We’re really looking forward to seeing how this research translates from the SMIB model out into real-world operation and see what kind of practical deployment looks like next.

Dev: Well, let's keep our eyes on these papers and get ready for whatever comes next in the queue.

Conclusion: Rosa: So we've just covered the paper "Flexible Training Workloads in Large-Scale AI Data Centers for Transient-Stability Support in Transmission Constrained Power Systems," and it seems they’ve shown a way to leverage AI data center flexibility for grid stability during disturbances.

Dev: I gotta say, Rosa, from my end as someone who deals with loop rates and latency, the concept of using training workloads as a fast corrective resource is actually compelling because it’s so much faster than traditional measures like ESS charging or dynamic braking resistors.

Taro: I agree with Dev; what interests me most is how the system performs when things go wrong in the real world; can this mechanism handle misbehaving loads or unexpected grid conditions?

Rosa: Well, the paper evaluates it across three different power systems—a small SMIB, a larger IEEE thirty-nine-bus system, and a Korean power system—and shows that TILS can increase the generation limit in all of them.

Dev: That's significant; seeing it work across such varied topologies is what gives me confidence about its applicability beyond just one specific grid configuration.

Taro: I wonder if this approach scales well when we move from an ideal step-increase model to the more realistic finite ramp-up and scheduling delays that you mentioned in the text.

Rosa: The authors acknowledged that they modeled an idealized step increase to isolate the core effects of response magnitude, timing, and location before addressing those more complex real-world dynamics.

Dev: That’s fair; their limitation is exactly that they didn't model the sequential workload activation or communication delays fully, but they did give us estimates for component timescales relevant to TILS activation.

Taro: So the next step for this research seems to be moving from idealized models to validating the complete end-to-end chain, including disturbance detection and actual workload verification at a multi-megawatt scale.

Rosa: Exactly; it’s a clear path forward, showing that TILS is not a replacement for reinforcement but an additional demand-side option when the right operational conditions are met.

Dev: It sounds like we have some solid groundwork here for how AI infrastructure could play a role in proactive stability support.

Taro: I'm definitely curious to see if this concept of using flexible workloads to actively influence generator acceleration becomes a standard consideration in future power system studies.

eess.SY, cs.SY

Submitted: 2026-08-31

Updated: 2026-09-25

Comments: 11 pages, 7 figures, 1 table

Code: https://github.com/pnnltestsystem/Enhanced-IEEE-39-Bus-System-with-Inv

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 79/100

The gist: The rapid expansion of large-scale artificial intelligence (AI) data centers is adding substantial, concentrated, and rapidly varying loads to transmission-constrained power systems.

Key concepts

Training-Induced Load Surge (TILS): A fast demand-side strategy that starts or resumes flexible AI training workloads after a fault clears to increase active power demand at electrically effective locations. This response is measured in terms of load (MW) and time (s).
Transient Stability Support: The ability of a power system to remain stable during short-lived disturbances, such as faults. The paper investigates how TILS can influence the active power response during this critical post-fault window.
Electrical Siting and Timing: The effectiveness of TILS depends on when the workload is activated (timing) and where it is located electrically (siting). Optimization aims for activation at electrically effective buses near generators with minimal delay, aiming for sub-second response times.
Complementary Resource: TILS should be viewed as a supplement to existing stability-enhancing measures rather than a replacement. Deployment requires sufficient electrical headroom, flexible workloads, and reliable grid-triggered activation before an event occurs.

Terminology

Artificial Intelligence (AI) — 5× in this paper · explained in 2 episodes
Verification — 3× in this paper · explained in 2 episodes
Core — 2× in this paper · explained in 2 episodes
Benchmark — 1× in this paper · explained in 2 episodes
Foundation Model — 1× in this paper · explained in 3 episodes
Foundation Models — 1× in this paper · explained in 7 episodes
Trace — 1× in this paper · explained in 2 episodes
Transformer — 1× in this paper · explained in 2 episodes

Summary

The paper proposes traininginduced load surge (TILS), a fast demand-side strategy that initiates or resumes flexible AI training workloads following fault clearing. Fig. 1(b) illustrates the overall concept and operating sequence of TILS. During the critical post-fault period, TILS rapidly increases AI datacenter active-power demand at electrically effective locations. Unlike conventional demand-response strategies that primarily reduce or shift electricity consumption, the proposed TILS repurposes upward load flexibility as a fast corrective resource, increasing local demand and thereby the electrical output of accelerating generators.

The proposed strategy is evaluated using three power systems of increasing complexity. First, a single-machine infinite-bus (SMIB) system is used to clarify the underlying physical mechanism and quantify the effect of TILS on the transient-stability-constrained generation limit. The analysis is then extended to the IEEE 39-bus system to examine the effectiveness of TILS in a multi-machine network and evaluate the effects of activation delay and electrical siting. Finally, TILS is assessed using a largescale Korean power system in which regional generation is operationally constrained to maintain transient stability [18]. Results from all case studies demonstrate that TILS can increase the transient-stability-constrained generation limit. Larger TILS responses, earlier activation, and siting at buses with a stronger electrical influence on the critical generators provide greater generation-limit increases, whereas delayed activation and electrically remote siting reduce the stabilizing effect.

The main contributions of this paper are summarized as follows:

• This paper repurposes rapid increases in AI datacenter demand, typically viewed as operational challenges, as a transient-stability support resource.

• TILS is proposed as a grid-triggered strategy that coordinates flexible AI training workloads at electrically effective locations following fault clearing to limit first-swing generator acceleration.

• The proposed TILS is evaluated in the SMIB, IEEE 39-bus, and Korean power systems, quantifying the effects of response magnitude, activation delay, and electrical siting.

The remainder of this paper is organized as follows. Section II reviews the relevant transient-stability fundamentals and introduces the proposed TILS strategy. Section III presents case-study results for the SMIB, IEEE 39bus, and Korean power systems. Section IV discusses the implications of TILS for power-system stability support, along with its deployment requirements, validation needs, and applicability. Section V concludes the paper.

In a single machine infinite-bus (SMIB) system under normal operating conditions, the generator operates close to steady state such that the mechanical input power approximately balances the electrical output power, i.e., Pm ≈ Pe. However, when a fault occurs in the SMIB system, the associated voltage depression reduces the power transferred through the transmission corridor, Ptrans, and consequently decreases the generator electrical output Pe. Because the mechanical input Pm changes negligibly over the first-swing timescale, the resulting imbalance Pm − Pe becomes positive. The generator therefore accelerates, causing its rotor angle to increase. If a fault is cleared by tripping one of the two transmission lines, the post-fault network is weaker than the pre-fault network because the equivalent transfer reactance Xeq increases. Consequently, the maximum power that can be transferred through the remaining transmission corridor is reduced even after fault clearing. If the post-fault electrical output remains below the mechanical input, positive accelerating power persists, thereby causing the rotor angle to continue increasing. Transient instability occurs when the post-fault system cannot provide sufficient decelerating power to limit the rotor-angle excursion, eventually causing the generator to lose synchronism with the grid. As indicated by (1), this requires a mismatch between the mechanical input power Pm and the electrical output power Pe to be reduced sufficiently rapidly. From (2), a rapid increase in Plocal provides a demand-side means of reducing this mismatch. This observation provides the physical basis for the proposed method.

The TILS strategy is modeled as an ideal step increase in AI data-center active-power demand, initiated after a prescribed activation delay following fault clearing. The activation delay is defined as an effective end-to-end delay from fault clearing to the establishment of the commanded TILS response, thereby representing the combined effects of disturbance detection, signal transmission, workload scheduling, and load activation. The step magnitude represents the aggregate electrical load increase produced by the activated AI training workloads and is varied to evaluate different TILS response magnitudes. Finite load ramp-up, staged server activation, and workload-level dynamics are not explicitly modeled. Instead, the case studies evaluate multiple activation delays to quantify the sensitivity of TILS performance to response timing. Because the additional local demand must be established during the critical first-swing interval, shorter activation delays more effectively suppress generator acceleration, whereas longer delays reduce the stabilizing effect of TILS.

The stability assessment follows the criteria defined in Section IIE, with time-domain simulations performed using PSS/E [22]. The system-specific performance metrics are defined as follows:

• SMIB system: The stability-constrained generation limit is defined as the maximum sending-end generation level at which the generator remains synchronized following the specified contingency. This limit is determined both without and with TILS, and the resulting increase in allowable generation is used to quantify the stabilizing effect of TILS. For selected target generation levels, the TILS magnitude required to maintain stability is also evaluated under different activation delays.

• 39-bus system: The stability-constrained generation limit is defined as the maximum output of the generator at Bus 32 that can be accommodated without loss of synchronism. This limit is re-evaluated with a 100 MW TILS response applied separately at each candidate location. The increase relative to the noTILS case is used to quantify the stabilizing effect of TILS and compare the effectiveness of the two siting configurations.

• Korean system: The eastern-area generation limit is defined as the maximum aggregate generation in the region for which synchronism is maintained following the specified 765-kV contingency. Aggregate generation in the eastern region is increased in discrete steps, while generation outside the region is reduced in accordance with the current Korean operating rule.

The results across all three systems demonstrate that TILS can increase the transient-stability-constrained generation limit. The magnitude of this increase depends on the TILS response magnitude, activation delay, and electrical siting. Earlier activation and siting at buses with a strong electrical influence on the critical generators consistently result in larger increases in the generation limit.

The paper concludes that TILS is not a replacement for transmission reinforcement or established corrective controls; rather, it provides an additional demand-side option when sufficient electrical headroom, flexible workloads, and reliable grid-triggered activation are available. Practical deployment will require system-specific dynamic studies and end-to-end validation of grid-triggered TILS responses at the multi-megawatt scale. The stabilizing mechanism is not specific to AI data centers; other large-scale flexible loads may provide similar transient-stability support if they can increase demand rapidly and reliably, maintain sufficient upward load-response capability, and are located at electrically effective buses. However, the benefits are expected primarily in systems where power transfer from a generation-rich area is limited by transient stability and the controllable load has a strong electrical influence on the accelerating generator group. Accordingly, systemspecific dynamic studies are required before deploying TILS or analogous load-based corrective controls.

TABLE I

Comparison of Conventional Measures for Transient Stability Enhancement and TILS

Option Dynamic braking resistor ESS charging Generator tripping Transmission reinforcement TILS

Activation or deployment timescale Approximately 100 ms. Requires dedicated equipment and additional capital investment. Several hundred ms. Requires storage and inverter capacity; costs can be substantial at the hundreds-of-megawatts scale; sufficient charging headroom must be available. Approximately 100 ms. Reduces available generation and may introduce frequency-stability risks if the generation loss is large. Years to decades. Requires major capital investment, long lead times, permitting, and public acceptance. Potentially subsecond with pre-staged workloads; end-to-end validation is required. Requires flexible workload headroom, grid-triggered control, and validation of rapid load activation.

The paper notes that the idealized step increase in active-power demand adopted in this study isolates the effects of response magnitude, activation delay, and electrical siting. Actual AI data-center responses may involve finite ramp rates, communication and scheduling delays, and sequential workload activation. Although large subsecond power variations have been observed in AI training workloads [4], [9], these observations do not yet demonstrate reliable, grid-triggered TILS activation at the multi-megawatt scale. Representative component timescales relevant to TILS activation are estimated based on prior studies and technical documentation [21], [32]–[34]. For pre-staged GPU activation, these timescales correspond to an estimated aggregate response time of approximately 0.16 s after fault clearing, suggesting that subsecond TILS activation could be feasible. However, this estimate is component-based rather than an end-to-end demonstration. Future gridtriggered, multi-megawatt demonstrations should therefore validate the complete response chain, from disturbance detection and signal transmission to workload activation and verification of the delivered active-power response.

The paper also highlights that AI training workloads may be particularly suitable for the proposed TILS for two reasons: their operational flexibility and their ability to utilize available power headroom in mixed-use AI data centers. First, training workloads can provide greater temporal and scheduling flexibility than inference workloads. In general, AI inference workloads arise from real-time user requests and are therefore subject to strict latency and service-quality requirements [35]. In contrast, AI model training is typically decoupled from real-time user requests, allowing adjustments in execution timing and scheduling [36], [37]. A recent study on demand response in AI data centers found that training workloads provided greater regulation flexibility than inference workloads, which was attributed to the longer and more malleable execution structure of training jobs [36], [37]. This capability has also been demonstrated through dynamic resource scaling and suspend–resume scheduling of realworld training jobs [37]. Second, AI training workloads can provide a practical means of utilizing available power headroom during inference operation without directly modulating latencysensitive inference services. Measurements from operational LLM clusters showed that inference clusters retained substantially greater power headroom than training clusters [9], indicating that inference operation may leave considerable capacity for additional electrical load. Furthermore, a GPU time-sharing system demonstrated that unused GPU capacity can be utilized by training workloads while preserving the performance of the primary inference service [35]. Therefore, the flexibility of AI training workloads and the available headroom for additional electrical load during inference operation support the use of training workloads as a controllable TILS resource in mixeduse AI data centers.

References

[1] T. Brown et al., “Language models are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020.

[2] J. Hoffmann et al., “Training compute-optimal large language models,” in Proceedings of the 36th International Conference on Neural Information Processing Systems, 2022, pp. 30 016–30 03016–30 031

[4] North American Electric Reliability Corporation, “Characteristics and risks of emerging large loads,” North American Electric Reliability Corporation, Report, Jul. 25

[9] P. Patel et al., “Characterizing power management opportunities for LLMs in the cloud,” in Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3, 2024, pp. 207–222.

[11] M.-S. Ko and H. Zhu, “Wide-area power system oscillations from large-scale AI workloads,” IEEE Transactions on Power Systems, pp. 1–14, 2026.

[35] Z. Bai, Z. Zhang, Y. Zhu, and X. Jin, “PipeSwitch: Fast pipelined context switching for deep learning applications,” in 14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20), 2020, pp. 499–514.

[36] F. Acun, C. Hankendi, E. Levine, H. Reynolds, J. Bardwick, and A. K. Coskun, “Investigating power consumption flexibility of AI data centers for demand response participation,” in Proceedings of the 17th ACM International Conference on Future and Sustainable Energy Systems, 2026, pp. 344–348.

[37] W. A. Hanafy, Q. Liang, N. Bashir, D. Irwin, and P. Shenoy, “Carbonscaler: Leveraging cloud workload elasticity for optimizing carbon-efficiency,” Proceedings of the ACM on Measurement and Analysis of Computing Systems, vol. 7, no 3 pp 1–28 2023.

[4] North American Electric Reliability Corporation, “Characteristics and risks of emerging large loads,” North American Electric Reliability Corporation, Report, Jul. 25

[10] Y. Li, M. Mughees, Y. Chen, and Y. R. Li, “The unseen AI disruptions for power grids: LLM-induced transients,” arXiv preprint arXiv:2409.11416, 2024

[18] D. R. Aryani, H. Song, and Y.-S. Cho, “Operation strategy of battery energy storage systems for stability improvement of the korean power system,” Journal of Energy Storage, vol. 56, p. 106091, 2022

[19] P. Kundur et al., “Definition and classification of power system stability ieee/cigre joint task force on stability terms and definitions,” IEEE Transactions on Power Systems, vol. 19, no 3 pp 1387–1401, Aug. 2004

[22] Program Operation Manual PSS® E 35.1.0, Siemens Power Technologies International, May 2020

[23] pnnltestsystem, “IEEE 39-Bus Test System,” https://github.com/pnnltestsystem/Enhanced-IEEE-39-Bus-System-with-Inv

erter-based-Resources-on Multi Time Scale Platforms, gitHub repository, accessed: 2026.

The paper is also titled Flexible Training Workloads in Large-Scale AI Data Centers for Transient Stability Support in Transmission Constrained Power Systems. The summary is as follows:

The rapid expansion of large-scale artificial intelligence (AI) data centers is adding substantial, concentrated, and rapidly varying loads to transmission-constrained power systems. Although such load variations are generally regarded as operational challenges, this paper presents an alternative perspective in which the upward load flexibility of AI data centers could be coordinated for transient-stability support. To this end, this paper proposes training-induced load surge (TILS), a fast demand-side strategy that initiates or resumes flexible AI training workloads after fault clearing to increase active-power demand at electrically effective locations. The resulting load increase allows accelerating generators to supply additional electrical power, thereby reducing the accelerating power imbalance and limiting the first-swing rotor-angle excursion. The underlying mechanism is first clarified in a single machine infinite-bus (SMIB) system and then evaluated in the IEEE 39-bus system and a large-scale Korean power system. Results across all three systems demonstrate that TILS can increase the transient-stability-constrained generation limit. Larger responses, earlier activation, and siting at buses with a stronger electrical influence on the critical generators provide greater generation-limit increases. These results suggest that the upward load-response capability of AI data centers can provide complementary transient-stability support when sufficient electrical headroom, flexible workloads, and reliable grid-triggered activation are available. The TILS strategy is evaluated using three power systems of increasing complexity: a single machine infinite-bus (SMIB) system to clarify the underlying physical mechanism and quantify the effect of TILS on the transient-stability-constrained generation limit; the IEEE 39-bus system to examine the effectiveness of TILS in a multi-machine network and evaluate the effects of activation delay and electrical siting; and a large-scale Korean power system in which regional generation is operationally constrained to maintain transient stability. The paper details that TILS is most effective when sufficient upward load-response capability is available at electrically effective locations and activated rapidly following a disturbance. The authors emphasize that TILS should be regarded as a complement to, rather than a replacement for, existing stability-enhancing measures. Deployment requirements include fast activation, sufficient upward load-response capability before a contingency occurs, and siting that reflects the electrical influence of the additional demand on the generators that dominate the post-fault instability. The paper concludes that TILS is not a replacement for transmission reinforcement or established corrective controls; rather, it provides an additional demand-side option when sufficient electrical headroom, flexible workloads, and reliable grid-triggered activation are available.

The TILS strategy is evaluated using three power systems of increasing complexity: a single machine infinite-bus (SMIB) system to clarify the underlying physical mechanism and quantify the effect of TILS on the transient-stability-constrained generation limit; the IEEE 39-bus system to examine the effectiveness of TILS in a multi-machine network and evaluate the effects of activation delay and electrical siting; and a large-scale Korean power system in which regional generation is operationally constrained to maintain transient stability. The paper details that TILS is most effective when sufficient upward load-response capability is available at electrically effective locations and activated rapidly following a disturbance. The authors emphasize that TILS should be regarded as a complement to, rather than a replacement for, existing stability-enhancing measures. Deployment requirements include fast activation, sufficient upward load-response capability before a contingency occurs, and siting that reflects the electrical influence of the additional demand on the generators that dominate the post-fault instability. The paper concludes that TILS is not a replacement for transmission reinforcement or established corrective controls; rather, it provides an additional demand-side option when sufficient electrical headroom, flexible workloads, and reliable grid-triggered activation are available.

TABLE I

Comparison of Conventional Measures for Transient Stability Enhancement and TILS

Option Dynamic braking resistor ESS charging Generator tripping Transmission reinforcement TILS

Improvements for AI systems

Based on the provided research paper, here are specific improvements that can be made to AI systems by implementing the proposed Training-Induced Load Surge (TILS) strategy, along with what these improved systems can achieve:

Improve AI Data Center Grid Interaction via TILS Implementation:
Enhance Transient Stability Support Post-Fault Clearing:
Optimize Load Response Timing and Siting for Maximum Stability Gain:
Enable Complementary Power System Resilience with Flexible Workloads:

Specific Capabilities of the Improved AI System (TILS-Enabled):

The improved system will act as a fast corrective resource by rapidly initiating or resuming flexible, pre-staged AI training workloads immediately following a grid disturbance (like a line fault). This increases local active power demand at electrically effective locations to directly boost the electrical output of nearby generators.
The system can provide immediate first-swing transient stability support by reducing the accelerating power imbalance between mechanical input and electrical output during the critical post-fault interval, thereby limiting or preventing rotor-angle excursions and maintaining synchronism in transmission-constrained power systems.
The improved system's deployment will be optimized by ensuring workloads are activated at electrically effective buses (near critical generators) with minimal activation delay (ideally sub-second), maximizing the generation limit increase achieved while minimizing the required load response magnitude.
The AI data center, when utilized as a TILS resource, can transition from being solely a consumer to an active participant in power system stability, providing a complementary corrective action alongside traditional measures like transmission reinforcement or energy storage systems (ESS).