Mitigation of Datacenter Demand Ramping and Fluctuation using Hybrid ESS and Supercapacitor

summary

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…

In short

The study proposes a hybrid energy storage system (BESS and supercapacitor) control framework to smooth power fluctuations in hyperscale AI datacenters. It uses coordinated multi-timescale control, including ramp-tracking and frequency stabilization, to effectively manage load variations caused by GPU training cycles, ensuring stable power delivery.

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 used across episodes

This episode discusses

The paper

Mitigation of Datacenter Demand Ramping and Fluctuation using Hybrid ESS and Supercapacitor · Read on arXiv

Chandra Family Department of Electrical & Computer Engineering at The University of Texas at Austin · Department of Electrical Engineering at Sangmyung University

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.

DOI: 10.1109/PESGM58988.2026.11693182

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.

More episodes

← Home