Mitigation of Datacenter Demand Ramping and Fluctuation using Hybrid ESS and Supercapacitor
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "Mitigation of Datacenter Demand Ramping and Fluctuation using Hybrid ESS and Supercapacitor".
Paper summary: Rosa: So we're talking about this paper now, "Mitigation of Datacenter Demand Ramping and Fluctuation using Hybrid ESS and Supercapacitor." Essentially, they’re proposing this hybrid energy storage system framework for smoothing out power variations in massive AI datacenters that have huge load swings.
Dev: Right, so the core idea is integrating a battery energy storage system, or BESS, with a supercapacitor—the SC—through some coordinated control strategy to handle these big load variations.
Taro: It seems like the main problem they’re tackling is these two distinct types of power issues: quasi-periodic fluctuations and abrupt ramping patterns that show up in AI training workloads.
Rosa: The paper sets out that the fluctuation part comes from cyclic variations around zero point, which is happening at a frequency of approximately zero point zero five hertz, and the ramp behavior shows full-range transitions between idle and peak operating conditions because of how much GPU utilization is going on > page two.
Dev: And to manage this, they’ve modeled the datacenter demand by taking those GPU level traces and making them synthetic for the aggregate demand, which includes some additional components not just from the raw GPU power consumption > page two.
Taro: I see they specifically generated a synthetic profile to show this behavior because real measurements alone weren't sufficient to capture everything about the datacenter load > page two.
Rosa: The paper claims that this hybrid system handles these issues by using the supercapacitor for rapid variations and transients, while the battery compensates for those slower deviations > page one <ref:2512.08076#pg1>.
Dev: They define a high pass filter, or HPF, to separate the demand into slow and fast parts, assigning them to the battery via a leaky integral controller and to the supercapacitor through a phase lead proportional derivative controller with some extra stuff > page one <ref:2512.08076#pg1,via a leaky integral controller and to the>.
Taro: I'm curious how they handle those sharp ramps or severe fluctuations because they note that the supercapacitor might have some response delay that hinders a timely reaction > page three <ref:2512.08076#pg1>.
Rosa: So to fix that, they add an extra ramp related term into the supercapacitor control loop, calling it P ref one SC equals GHPF of s delta of s plus omega ramp T effective R hat DC of s > page three <ref:2512.08076#pg1>.
Paper summary: Dev: That ramp term depends on an estimated ramp rate called R hat DC, which they figure out using a discrete time Kalman filter or KF state vector that includes the slow varying baseline and the rate of change > page three <ref:2512.08076#pg1>.
Taro: The paper mentions an adaptive scaling factor omega ramp that uses a sigmoid function to activate this ramp term only when the estimated ramp rate exceeds some reference value s ref > page three <ref:2512.08076#pg1>.
Rosa: And then there’s a dynamic supercapacitor command manager introduced to make sure both the battery and the supercapacitor keep a sustainable state of charge trajectory over time > page one <ref:2512.08076#pg1>.
Dev: The BESS controller on the other hand uses an integral controller and a repetitive control filter, which they use to actively reject repeating frequency components that might cause oscillations between the two storage units > page three <ref:2512.08076#pg1>.
Taro: It sounds like they’re using these different control structures—PD for the SC, I and RC for the BESS—to keep everything balanced in terms of both power delivery and system frequency stability > page one <ref:2512.08076#pg1>.
Rosa: And from what I see, the experimental results on their synthetic one gigawatt system show that both the battery and supercapacitor outputs follow their command trajectories really closely > page 6a <ref:2512.08076#pg1>.
Dev: They say this scheme ensures sustainable operation and long term stability even when there are prolonged and stochastic load variations, keeping the grid-supplied power nearly constant and minimizing frequency deviations around sixty hertz > page 6a <ref:2512.08076#pg1>.
Taro: So, for someone just listening to the show, what does this actually mean for their day? It means that even if you’re powering an AI training job that has crazy sudden power demands, the system is designed to keep things stable without big dips in electricity or frequency wobbles > page 6a <ref:2512.08076#pg1>.
Rosa: Exactly. The paper shows how this hybrid approach, with all those feedforward delays and ramp tracking terms, can substantially reduce the power and frequency deviations compared to just running them separately > page one <ref:2512.08076#pg1>.
Paper summary: Dev: But they also have to be careful about what this setup can't do, because they flag that the supercapacitor might still exhibit some relative response delay when facing severe ramping or sharp fluctuations > page three <ref:2512.08076#pg1>.
Taro: So it’s not a perfect solution where everything reacts instantly to every single spike, which is a key limitation they point out > page three <ref:2512.08076#pg1>.
Rosa: They also mentioned that the entire system relies on their synthetic modeling of the AI datacenter load to generate these specific test conditions > page two.
Dev: So the implication is that this control framework gives engineers a concrete way to design storage for these very specific types of massive power demands, moving beyond just simple response times > page one <ref:2512.08076#pg1>.
Taro: It’s interesting how they use the Kalman filter to estimate the ramp rate because it allows them to adapt the ramp compensation term based on what’s actually happening in real time > page three <ref:2512.08076#pg1>.
Rosa: So, putting it simply, this paper is about using a battery and a supercapacitor together with very specific, adaptive control commands to smooth out the rough power demands of huge AI datacenters > page one <ref:2512.08076#pg1>.
Dev: The title of the paper is "Mitigation of Datacenter Demand Ramping and Fluctuation using Hybrid ESS and Supercapacitor," written by Min-Seung Ko and Hao Zhu > page one <ref:2512.08076#pg1,Mitigation of Datacenter Demand Ramping and Fluctuation using Hybrid ESS and Supercapacitor>.
Taro: It suggests that for real world applications, especially with those massive workloads, having a coordinated multi timescale control system is necessary to manage the different dynamic regimes of the load > page one <ref:2512.08076#pg1>.
Rosa: It’s about showing how you can use feedforward delay compensation and ramp tracking terms to enhance performance without violating power or state of charge constraints > page one <ref:2512.08076#pg1>.
Dev: The results confirm that under prolonged and stochastic load variations, this proposed HESS substantially reduces power and frequency deviations compared to standalone operation > page 6a <ref:2512.08076#pg1>.
Taro: What this means for the future is that we need these coordinated control methods when dealing with large digital loads because they create these complex dynamic stresses on the grid > page one <ref:2512.08076#pg1>.
Conclusion: Rosa: So we're wrapping up this look at the paper "Mitigation of Datacenter Demand Ramping and Fluctuation using Hybrid ESS and Supercapacitor."
Dev: The authors are Ko and Zhu, they’re tackling power smoothing for massive AI datacenters by putting a battery energy storage system and a supercapacitor together with some clever control.
Rosa: It really boils down to coordinating these two different types of storage units—the slow battery and the fast supercapacitor—to keep the power supply steady when the AI training jobs suddenly ramp up or fluctuate.
Dev: They use this hybrid control framework, incorporating things like feedforward delay compensation and specific ramp tracking terms to manage those rapid changes without messing up the overall system.
Rosa: The results on their synthetic one gigawatt system show that both storage units follow their designed commands very closely, which means they keep the grid power almost constant and frequency stays steady around sixty hertz even when the load swings wildly.
Dev: That's a solid outcome for stability, but we gotta remember they did this using a synthetic model of AI training workloads, so we need to watch how it holds up when the real world gets messy.
Rosa: Exactly. The paper shows that combining the right control logic—like using a Kalman filter to estimate ramp rates—actually makes the system perform better than just running one storage type alone.
Dev: It’s a practical approach for engineers trying to design reliable power systems for these huge computational loads, but we still need to figure out the safety case when things get truly unpredictable.
Chandra Family Department of Electrical & Computer Engineering at The University of Texas at Austin · Department of Electrical Engineering at Sangmyung University
eess.SY, cs.SY
Submitted: 2025-12-08
Updated: 2025-12-08
DOI: 10.1109/PESGM58988.2026.11693182
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 93/100
The gist: This paper proposes a hybrid energy storage system (HESS)-based control framework that enables comprehensive power smoothing for hyperscale AI datacenters with large load variations The proposed HESS…
Key concepts
- Hybrid Energy Storage System (HESS)
- This system combines a battery energy storage system (BESS) and a supercapacitor (SC). The BESS handles slower power deviations, while the SC manages very fast spikes and rapid transients. They work together to provide comprehensive power smoothing for the datacenter load.
- Ramp Tracking
- Since AI workloads cause sudden changes in power demand (ramping), this technique is used to proactively adjust the storage system's output. It estimates how quickly the load is changing and applies a corrective signal to prevent large, sudden jumps in power consumption.
- Repetitive Control (RC) Filter
- This control component actively filters out repeating frequency components in the system. In this context, it helps mitigate oscillations that can occur between the BESS and SC when they are interacting with each other during rapid load changes.
Terminology
Summary
This paper proposes a hybrid energy storage system (HESS)-based control framework that enables comprehensive power smoothing for hyperscale AI datacenters with large load variations The proposed HESS integrates a battery energy storage system (BESS) and a supercapacitor (SC) through coordinated multi-timescale control Case studies verify that the proposed method effectively suppresses both ramping and fluctuations, stabilizes the system frequency, and maintains sustainable state-of-charge (SoC) trajectories for both ESS and SC under prolonged, stochastic training cycles
System Modeling of AI Datacenters
The power fluctuations induced by AI training workloads exhibit both quasi-periodic fluctuations and abrupt ramping patterns The fluctuation component is primarily driven by cyclic variations with approximately 0.05 Hz, arising from the iterative gradient computation inherent to the training process In terms of ramp behavior, the measured trace clearly demonstrates full-range transitions between idle and peak operating conditions due to the high GPU utilization The synthetic datacenter load demand retains the dominant fluctuation with some modifications to reflect the aggregated behavior
Power Smoothing via HESS Control Designs
The goal of proposing to co-locate HESS in datacenters is to attain power smoothing of PDC(t) by utilizing the two heterogeneous and complementary resources The SC responds to the rapid variations and transient spikes via a command signal defined by P ref0 SC = GHPF(s)∆(s) = s/(s + 2πfc)∆(s) The BESS compensates for slower deviations by using a command signal defined by P ref0 ESS = (1 − GHPF(s)) ∆(s) To address severe ramping or sharp fluctuations, an additional ramp-related term is incorporated into the SC control loop as P ref1 SC = GHPF(s)∆(s) + ωrampTeffRˆDC(s)
Where the gains come from
The estimation of the ramp rate RˆDC is performed using a discrete-time Kalman filter (KF) The KF state vector is defined as z KFk = [∆¯, RDC⊤] An adaptive update for the parameter ωramp = γramp · γjerk is designed using a Sigmoid function This ramp term is activated only when the estimated ramp exceeds s ref The dynamic SoC command manager is introduced to ensure the long-term SoC balance of both BESS and SC
Design of BESS and SC Controllers
The SC controller adopts a proportional–derivative (PD) controller to provide phase lead compensation The enhanced SC control structure consists of three components: PD controller, feedforward (FF), and ramp tracking (RK) The RK term is used to address any remaining mismatch between the desired and actual ramp profiles as URK(s) = kRK RˆDC(s) − Rˆ SC(s) Similarly, the BESS controller consists of an integral (I) controller and a repetitive control (RC) filter The RC filter actively rejects repeating frequency components to mitigate interaction-induced oscillations between the SC and BESS
Experimental Settings and Results
The experiment was conducted on a synthetic simplified 1 GW system with specific parameters for the grid and HESS The frequency separation enables the HESS to handle distinct dynamic regimes of the datacenter load Both BESS and SC closely follow their command trajectories, maintaining the grid-supplied power nearly constant and minimizing frequency deviations The resulting responses of the BESS and SC outputs to their respective command signals are depicted in Fig. 6(a) Simulation results confirm that the proposed scheme ensures sustainable operation and long-term stability even under prolonged and stochastic load variations
Conclusion
This work presents a coordinated hybrid control framework for power smoothing of AI datacenter loads using co-located BESS and SC units The inclusion of feedforward delay compensation, ramp-tracking, and repetitive control terms further enhances dynamic performance without violating power or SoC constraints Simulation results confirm that the proposed HESS substantially reduces power and frequency deviations compared with standalone operation The system frequency remains tightly regulated around the nominal 60 Hz even under prolonged and stochastic load variations
REFERENCES
[1] NERC Large Load Task Force, “Characteristics and risks of emerging large loads,” North American Electric Reliability Corporation (NERC), Tech. Rep., 2025
[2] J. Aljbour, T. p. 4 REFERENCES
[1] NERC Large Load Task Force, “Characteristics and risks of emerging
large loads,” North American Electric Reliability Corporation (NERC),
Tech. Aljbour, T. Aljbour, T. Wilson, and P. Wilson, and P. Patel, “Powering intelligence: Analyzing
artificial intelligence and data center energy consumption,” EPRI White
Paper no. Patel, “Powering intelligence: Analyzing
artificial intelligence and data center energy consumption,” EPRI White
Paper no. 3002028905, Tech.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[3] M.-S. Ko and H. Zhu, “Wide-area power system oscillations from largescale AI workloads,” arXiv preprint arXiv:2508. p. 4 3002028905, Tech. Zhu, “Wide-area power system oscillations from largescale AI workloads,” arXiv preprint arXiv:2508.Where the paper says it
[4] E. Choukse et al. p. 5 Choukse et al. , “Power stabilization for AI training datacenters,” arXiv
preprint arXiv:2508. Choukse et al. , “Power stabilization for AI training datacenters,” arXiv
preprint arXiv:2508.Where the paper says it
Where the paper says it
[5] S. Kundu, K. Chatterjee, R. p. 5 14318, 2025. Chatterjee, R. Chatterjee, R. Hossain, S. Hossain, S. Nandanoori, and V. Nandanoori, and V. Adetola,
“Managing risks from large digital loads using coordinated grid-forming
storage network,” arXiv preprint arXiv:2508.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[6] V. T. Nguyen and J. p. 5 11080, 2025. Nguyen and J. Nguyen and J. Shim, “Virtual capacity of hybrid energy storage
systems using adaptive state of charge range control for smoothing
renewable intermittency,” IEEE Access, vol.Where the paper says it
Where the paper says it
[7] J. Xiao, P. Wang, and L. p. 5 126 951–126 964,
2020. Wang, and L. Wang, and L. Setyawan, “Hierarchical control of hybrid energy
storage system in DC microgrids,” IEEE Transactions on Industrial
Electronics, vol.Where the paper says it
Where the paper says it
[8] F. Orderud, “Comparison of Kalman filter estimation approaches for state space models with nonlinear measurements,” in Proc. p. 5 4915–4924, 2015. Orderud, “Comparison of Kalman filter estimation approaches for state
space models with nonlinear measurements,” in Proc. Orderud, “Comparison of Kalman filter estimation approaches for state
space models with nonlinear measurements,” in Proc. of Scandinavian
Conference on Simulation and Modeling, 2005, pp.Where the paper says it
Where the paper says it
[9] T. Chmielewski, W. p. 5 of Scandinavian
Conference on Simulation and Modeling, 2005, pp. Chmielewski, W. Chmielewski, W. Jarzyna, D. Jarzyna, D. Zielinski, K. Zielinski, K. Gopakumar, and ´
M. Gopakumar, and ´
M. Chmielewska, “Modified repetitive control based on comb filters for
harmonics control in grid-connected applications,” Electric Power Systems
Research, vol.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive HESS Control Framework: The system can effectively suppress both ramping and fluctuations by using
frequency-based power decomposition and a look-ahead compensation approach to systematically mitigate load ramps in addition to fluctuations.
-
SC Control Enhancements: The supercapacitor control loop can be improved with an additional ramp-related term, Pref1SC = GHPF(s)∆(s) + ωrampTeffRˆDC(s), which uses an estimated ramp rate with an adaptive scaling factor omega ramp to provide a more adequate reference value for ramp conditions.
-
Predictive Feedforward: A Kalman filter observer can be used to estimate the instantaneous slope and jerk in the demand profile, enabling a
predictive feedforward that prepares the SC ahead of steep ramps and improves transient tracking beyond pure feedback control.
-
Dynamic SoC Management: The system can maintain sustainable operation during long-duration training by incorporating a dynamic SoC command manager, which applies baseline correction of the desired demand and uses an idle interval to mitigate energy imbalance.
-
Coordinated Control Design: The BESS controller can use a leaky integral controller to prevent windup issues while maintaining a bounded control effort, and the SC controller can employ a proportional minus derivative (PD) controller with feedforward and ramp tracking components to achieve coordinated control across multiple frequency bands.
Abstract
This paper proposes a hybrid energy storage system (HESS)-based control framework that enables comprehensive power smoothing for hyperscale AI datacenters with large load variations. Datacenters impose severe ramping and fluctuation-induced stresses on the grid frequency and voltage stability. To mitigate such disturbances, the proposed HESS integrates a battery energy storage system (BESS) and a supercapacitor (SC) through coordinated multi-timescale control. A high-pass filter (HPF) separates the datacenter demand into slow and fast components, allocating them respectively to the ESS via a leaky-integral controller and to the SC via a phase-lead proportional-derivative controller enhanced with feedforward and ramp-tracking compensation. Adaptive weighting and repetitive control mechanisms further improve transient and periodic responses. Case studies verify that the proposed method effectively suppresses both ramping and fluctuations, stabilizes the system frequency, and maintains sustainable state-of-charge (SoC) trajectories for both ESS and SC under prolonged, stochastic training cycles.
Sources
- Wide-Area Power System Oscillations from Large-Scale AI Workloads
- Power Stabilization for AI Training Datacenters
- Managing Risks from Large Digital Loads Using Coordinated Grid-Forming Storage Network
Related papers
- Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems
- Exact and Approximate Convex Reformulation of Linear Stochastic Optimal Control with Chance Constraints
- Certified Set Convergence for Piecewise Affine Systems
- Symbolic Control for Stochastic Systems via Finite Parity Games
- Multiobjective optimization-based design and dispatch of islanded, hybrid microgrids for remote, off-grid communities in sub-Saharan Africa
- Physiology-informed layered sensing for intelligent human-exoskeleton interaction