betaFeedback

WorksEngineering

Grid-Mode-Aware AI Data Center Power Smoothing via Bayesian-Optimized Hybrid Energy Storage System

Abstract

AI data centers exhibit highly stochastic, volatile power demands from intensive workloads and synchronized GPU training cycles, which can excite grid natural frequencies and threaten system stability. This paper proposes a real-time, forecasting-independent control framework utilizing a Hybrid Energy Storage System (HESS) that coordinates a Battery Energy Storage System (BESS) and a supercapacitor (SC). Employing a parameterized frequency-decoupling and active modal-damping policy, the framework mitigates critical oscillatory load components while smoothing overall power fluctuations. To bypass the heavy computational burden of online receding-horizon optimization, this work introduces an offline Bayesian Optimization (BO) approach to optimally tune real-time control parameters, minimizing a multi-objective cost function that balances grid compliance penalties and storage degradation. Numerical simulations demonstrate that the proposed method ensures grid stability and strong target envelope adherence. Specifically, it reshapes extreme 75-160 MW raw demand fluctuations to largely fit within a tight 10 MW target band, with only brief transient excursions, while attenuating the critical 20 MW oscillatory component below 1 MW to suppress resonant excitation.

Introduction

The rapid growth of artificial intelligence (AI) is increasing both the number and the size of data centers (DCs). According to recent reports, more than 11,000 DCs were operating worldwide by early 2024, and their electricity consumption has become a noticeable part of the total power demand. In the United States, server electricity use increased from about 30 TWh in 2014 to nearly 100 TWh in 2023, mainly due to the rapid growth of AI computing infrastructure [1], [2].

Besides their high energy consumption, AI DCs can also create large and fast changes in power demand. During AI training, a large number of GPUs may switch together among computation, communication, and checkpointing stages, resulting in repeated and highly variable power fluctuations. These variations can make real-time power balancing and frequency regulation more difficult. More importantly, periodic or quasi-periodic load variations can continuously excite a poorly damped grid mode when their frequency is close to the natural frequency of the power system [2], [3]. Therefore, reducing only the overall load variation may not be enough. The frequency content of the load should also be considered so that components close to vulnerable grid modes can be specifically reduced. optimization, or short-term load forecast is required.

The main contribution of this paper is an offline BO-based grid-mode-aware HESS control framework for AI DC power smoothing. The proposed framework keeps explicit awareness of vulnerable grid modes while moving the computationally intensive parameter search to the offline stage, leading to a simple online control implementation without receding-horizon optimization. Numerical studies are used to evaluate its power-smoothing and grid-frequency performance and to examine the benefit of explicitly considering modal components in the controller design.

Proposed Methodology

The proposed framework coordinates a BESS and a SC to smooth the highly variable power demand of AI data centers and mitigate load components that may excite poorly damped grid modes, as shown in Figure 1. Instead of solving an online receding-horizon optimization problem, a parameterized frequency-decoupling and mode-aware control policy is executed at each sampling instant. Its control parameters are optimized offline using BO, thereby retaining low online computational complexity while accounting for grid-side performance and HESS operating costs.

Figure 1. Proposed BO-based grid-mode-aware HESS control architecture for AI data-center load smoothing and modal suppression.

AI Data Center Load and Grid Interaction Modeling

The total AI data center power demand, Pdc(t)P_{dc}(t), consists of the computational workload Pwork(t)P_{work}(t), auxiliary demand PauxP_{aux}, and an oscillatory component Posc(t)P_{osc}(t):

Pdc(t)=Pwork(t)+Paux+Posc(t).(1)P_{dc}(t) = P_{work}(t) + P_{aux} + P_{osc}(t). \tag*{(1)}

The computational workload is represented by an Ornstein–Uhlenbeck (OU) mean-reverting stochastic process augmented with compound Poisson burst events [9]. The utilization state η(t)\eta(t) is modeled as

τηη˙=−(η−μη)+ξ(t)+∑jAjδ(t−tj),(2)\tau_{\eta}\dot{\eta} = -(\eta-\mu_{\eta})+\xi(t)+\sum_j A_j\delta(t-t_j), \tag*{(2)}

where μη\mu_{\eta} denotes the nominal utilization, ξ(t)\xi(t) represents continuous stochastic variations, and the final term captures burst events occurring at Poisson arrival times tjt_j with random amplitudes AjA_j. The corresponding workload power is

Pwork(t)=pbase+η(t)(pfull−pbase).(3)P_{work}(t) = p_{base}+\eta(t)(p_{full}-p_{base}). \tag*{(3)}

Parameterized Real-Time HESS Control

An oscillatory component

Posc(t)=Aoscsin⁡(ωnat)(4)P_{osc}(t) = A_{osc}\sin(\omega_n a t) \tag*{(4)}

is superimposed to represent synchronized computational activity capable of exciting a vulnerable grid mode.

The net power observed by the grid is

Pg(t)=Pdc(t)−Pbat(t)−Psc(t),(5)P_g(t) = P_{dc}(t)-P_{bat}(t)-P_{sc}(t), \tag*{(5)}

where positive storage power denotes discharge toward the data center.

To identify grid-sensitive oscillatory content, Pg(t)P_g(t) is passed through a second-order band-pass filter centered at the vulnerable modal frequency [8]:

x˙m=Amcxm+BmcPg,(6)\dot{x}_m = A_m^c x_m+B_m^c P_g, \tag*{(6)}
zm=Cmcxm,(7)z_m = C_m^c x_m, \tag*{(7)}

where xmx_m denotes the filter state and zmz_m is the extracted modal component. The continuous modal dynamics and converter tracking dynamics are discretized using an exact Zero-Order Hold transformation.

The controller is parameterized by

θ=[τlpf,αsc,Kmode,Ksocbat,Ksocsc],(8)\theta= [\tau_{lpf},\alpha_{sc},K_{mode},K_{soc}^{bat},K_{soc}^{sc}], \tag*{(8)}

where τlpf\tau_{lpf} determines the time-scale separation, αsc\alpha_{sc} controls transient power sharing, KmodeK_{mode} scales modal compensation, and KsocbatK_{soc}^{bat} and KsocscK_{soc}^{sc} regulate the storage energy states.

  1. Frequency Decoupling: The slowly varying component of the data center load is extracted using a first-order low-pass filter,

Plpf,k=ρfPlpf,k−1+(1−ρf)Pdc,k,(9)P_{lpf,k} = \rho_f P_{lpf,k-1}+(1-\rho_f)P_{dc,k}, \tag*{(9)}

where

ρf=exp⁡(−Δtτlpf).(10)\rho_f = \exp\left(-\frac{\Delta t}{\tau_{lpf}}\right). \tag*{(10)}

The complementary transient component is

Ptr,k=Pdc,k−Plpf,k.(11)P_{tr,k} = P_{dc,k}-P_{lpf,k}. \tag*{(11)}

This decomposition enables fast load variations to be preferentially assigned to the SC while slower energy-balancing requirements are handled primarily by the BESS.

  1. Energy-State Restoration and Power Allocation: To prevent gradual drift of the storage energy states, proportional restoration terms are defined as

Psoc,kbat=Ksocbat(Ebat,0−Ebat,k),(12)P_{soc,k}^{bat} = K_{soc}^{bat}(E_{bat,0}-E_{bat,k}), \tag*{(12)}
Psoc,ksc=Ksocsc(Esc,0−Esc,k).(13)P_{soc,k}^{sc} = K_{soc}^{sc}(E_{sc,0}-E_{sc,k}). \tag*{(13)}

The SC power reference is

Psc,kref=αscPtr,k+Kmodezm,k+Psoc,ksc,(14)P_{sc,k}^{ref} = \alpha_{sc}P_{tr,k}+K_{mode}z_{m,k}+P_{soc,k}^{sc}, \tag*{(14)}

where the first term assigns a fraction of the transient component to the SC and the second introduces mode-aware compensation.

The BESS reference is

Pbat,kref=(1−αsc)Ptr,k+(Plpf,k−Pcenter)+Psoc,kbat,(15)P_{bat,k}^{ref} = (1-\alpha_{sc})P_{tr,k}+(P_{lpf,k}-P_{center})+P_{soc,k}^{bat}, \tag*{(15)}

where PcenterP_{center} denotes the desired center of the grid-side operating range. Thus, the BESS absorbs the remaining transient component while also correcting slower deviations of the grid demand from its reference level.

The power references are first restricted by the rated converter limits. The finite converter response is represented using first-order tracking dynamics,

P~i,kact=ρiP~i,k−1act+(1−ρi)Pi,kref,i∈{bat,sc},(16)\widetilde{P}^{\mathrm{act}}_{i,k}=\rho_i\widetilde{P}^{\mathrm{act}}_{i,k-1}+(1-\rho_i)P^{\mathrm{ref}}_{i,k},\qquad i\in\{bat,sc\}, \tag*{(16)}

with

ρi=exp⁡(−Δtτi).(17)\rho_i=\exp\left(-\frac{\Delta t}{\tau_i}\right). \tag*{(17)}

Energy-Feasible Storage Operation

In addition to the rated converter limits, the commanded charge and discharge powers are constrained by the energy currently available in each storage device. Let

Δth=Δt3600(18)\Delta t_h=\frac{\Delta t}{3600} \tag*{(18)}

denote the sampling interval expressed in hours. The energy-dependent discharge and charging limits are

Pi,kdis,E=(Ei,k−E‾i)ηiΔth,(19)P^{\mathrm{dis},E}_{i,k}=\frac{(E_{i,k}-\underline{E}_i)\eta_i}{\Delta t_h}, \tag*{(19)}
Pi,kch,E=E‾i−Ei,kηiΔth.(20)P^{\mathrm{ch},E}_{i,k}=\frac{\overline{E}_i-E_{i,k}}{\eta_i\Delta t_h}. \tag*{(20)}

The feasible instantaneous limits are therefore

P‾i,kdis=min⁡(P‾i,Pi,kdis,E),(21)\overline{P}^{\mathrm{dis}}_{i,k}=\min\left(\overline{P}_i,P^{\mathrm{dis},E}_{i,k}\right), \tag*{(21)}
P‾i,kch=min⁡(P‾i,Pi,kch,E),(22)\overline{P}^{\mathrm{ch}}_{i,k}=\min\left(\overline{P}_i,P^{\mathrm{ch},E}_{i,k}\right), \tag*{(22)}

and the actual power is obtained by projecting the converter output onto the feasible interval:

Pi,kact=Π[−P‾i,kch,P‾i,kdis](P~i,kact).(23)P^{\mathrm{act}}_{i,k}=\Pi_{[-\overline{P}^{\mathrm{ch}}_{i,k},\overline{P}^{\mathrm{dis}}_{i,k}]}\left(\widetilde{P}^{\mathrm{act}}_{i,k}\right). \tag*{(23)}

This prevents either storage device from delivering power below its minimum energy level or absorbing power above its maximum energy level.

The actual power is separated into nonnegative charging and discharging components,

Pi,kdis=max⁡(0,Pi,kact),(24)P^{\mathrm{dis}}_{i,k}=\max\left(0,P^{\mathrm{act}}_{i,k}\right), \tag*{(24)}
Pi,kch=max⁡(0,−Pi,kact),(25)P^{\mathrm{ch}}_{i,k}=\max\left(0,-P^{\mathrm{act}}_{i,k}\right), \tag*{(25)}

which inherently prevents simultaneous charging and discharging.

The energy dynamics are then

Ei,k+1=Ei,k+Δth(ηiPi,kch−1ηiPi,kdis),(26)E_{i,k+1}=E_{i,k}+\Delta t_h\left(\eta_iP^{\mathrm{ch}}_{i,k}-\frac{1}{\eta_i}P^{\mathrm{dis}}_{i,k}\right), \tag*{(26)}

subject to

E‾i≤Ei,k≤E‾i.(27)\underline{E}_i\le E_{i,k}\le\overline{E}_i. \tag*{(27)}

Grid and HESS Performance Objective

For a given parameter vector θ\boldsymbol{\theta} and workload realization, the controller is evaluated using a cumulative objective combining grid-side performance and HESS operating costs:

J(θ)=∑k=1NΔth(Jgrid,k+Jhess,k).(28)J(\boldsymbol{\theta})=\sum_{k=1}^{N}\Delta t_h\left(J_{\mathrm{grid},k}+J_{\mathrm{hess},k}\right). \tag*{(28)}

The grid-side term penalizes violations of the desired power envelope, grid ramp-rate limit, and modal-power limit:

Jgrid,k=cpϵp,k2+crϵr,k2+cmϵm,k2,(29)J_{\mathrm{grid},k}=c_p\epsilon_{p,k}^{2}+c_r\epsilon_{r,k}^{2}+c_m\epsilon_{m,k}^{2}, \tag*{(29)}

where

ϵp,k=max⁡(0,P‾g−Pg,k,Pg,k−P‾g),(30)\epsilon_{p,k}=\max\left(0,\underline{P}_g-P_{g,k},P_{g,k}-\overline{P}_g\right), \tag*{(30)}
ϵr,k=max⁡(0,∣Pg,k−Pg,k−1Δt∣−R‾g),(31)\epsilon_{r,k}=\max\left(0,\left|\frac{P_{g,k}-P_{g,k-1}}{\Delta t}\right|-\overline{R}_g\right), \tag*{(31)}

and

ϵm,k=max⁡(0,∣zm,k∣−z‾m).(32)\epsilon_{m,k}=\max\left(0,\left|z_{m,k}\right|-\overline{z}_m\right). \tag*{(32)}

These quantities are treated as soft grid-side requirements. Consequently, temporary violations remain feasible but are penalized according to their magnitude.

The HESS-side cost consists of degradation and power-ramping terms:

Jhess,k=Jhess,kdeg+Jhess,kramp.(33)J_{\mathrm{hess},k}=J^{\mathrm{deg}}_{\mathrm{hess},k}+J^{\mathrm{ramp}}_{\mathrm{hess},k}. \tag*{(33)}

The degradation term penalizes storage energy throughput,

Jhess,kdeg=cbatdeg(ηbatPbat,kch+Pbat,kdisηbat)+cscdeg(ηscPsc,kch+Psc,kdisηsc),(34)J^{\mathrm{deg}}_{\mathrm{hess},k}=c^{\mathrm{deg}}_{\mathrm{bat}}\left(\eta_{\mathrm{bat}}P^{\mathrm{ch}}_{\mathrm{bat},k}+\frac{P^{\mathrm{dis}}_{\mathrm{bat},k}}{\eta_{\mathrm{bat}}}\right)+c^{\mathrm{deg}}_{\mathrm{sc}}\left(\eta_{\mathrm{sc}}P^{\mathrm{ch}}_{\mathrm{sc},k}+\frac{P^{\mathrm{dis}}_{\mathrm{sc},k}}{\eta_{\mathrm{sc}}}\right), \tag*{(34)}

while the ramping term is

Jhess,kramp=cbatΔ(Pbat,kact−Pbat,k−1actΔt)2+cscΔ(Psc,kact−Psc,k−1actΔt)2.(35)J^{\mathrm{ramp}}_{\mathrm{hess},k}=c^{\Delta}_{\mathrm{bat}}\left(\frac{P^{\mathrm{act}}_{\mathrm{bat},k}-P^{\mathrm{act}}_{\mathrm{bat},k-1}}{\Delta t}\right)^2+c^{\Delta}_{\mathrm{sc}}\left(\frac{P^{\mathrm{act}}_{\mathrm{sc},k}-P^{\mathrm{act}}_{\mathrm{sc},k-1}}{\Delta t}\right)^2. \tag*{(35)}

The combination of these terms encourages grid-compatible power trajectories while limiting unnecessary storage throughput and rapid cycling.

Offline Tuning via Bayesian Optimization

The relationship between θ\boldsymbol{\theta} and the cumulative closed-loop cost is nonlinear because of the stochastic workload, modal dynamics, converter response, saturation limits, energy constraints, and nonlinear penalty terms. The controller tuning problem is therefore treated as a bounded black-box optimization problem:

θ∗=arg⁡min⁡θ∈Θ E[J(θ;Pdc)].(36)\boldsymbol{\theta}^{*}=\underset{\boldsymbol{\theta}\in\Theta}{\arg\min}\ \mathbb{E}\left[J\left(\boldsymbol{\theta};P_{\mathrm{dc}}\right)\right]. \tag*{(36)}

To approximate the expectation over stochastic AI workloads, each candidate controller is evaluated on a fixed collection of independently generated training profiles. The BO objective is the corresponding sample-average cost:

J^train(θ)=1Ns∑s=1NsJ(θ;Pdc(s)).(37)\widehat{J}_{\mathrm{train}}(\boldsymbol{\theta})=\frac{1}{N_s}\sum_{s=1}^{N_s}J\left(\boldsymbol{\theta};P_{\mathrm{dc}}^{(s)}\right). \tag*{(37)}

Using the same workload realizations for all candidate controllers provides a consistent comparison between parameter sets and avoids introducing additional Monte-Carlo variation during the BO iterations.

A Gaussian Process (GP) is used as a probabilistic surrogate model of the objective landscape. The controller parameters are normalized before fitting the GP, while the observed objective values are standardized to improve numerical conditioning. Since the implementation maximizes the negative objective, the surrogate is defined for

y(θ)=−J^train(θ),(38)y(\boldsymbol{\theta})=-\widehat{J}_{\mathrm{train}}(\boldsymbol{\theta}), \tag*{(38)}

with

y(θ)∼GP(m(θ),k(θ,θ′)).(39)y(\boldsymbol{\theta})\sim\mathcal{GP}\left(m(\boldsymbol{\theta}),k(\boldsymbol{\theta},\boldsymbol{\theta}')\right). \tag*{(39)}

Following an initial set of controller evaluations, the logarithmic Expected Improvement (LogEI) acquisition function is used to identify the next parameter vector:

θnext=arg⁡max⁡θ∈Θ aLogEI(θ).(40)\theta_{\mathrm{next}} = \underset{\theta\in\Theta}{\arg\max}\ a_{\mathrm{LogEI}}(\theta). \tag*{(40)}

After each new evaluation, the observation is appended to the training data and the GP surrogate is refitted. The final controller corresponds to the evaluated parameter vector achieving the lowest sample-average training cost.

To assess generalization, the selected controller is subsequently evaluated on an independent set of stochastic workload realizations that are excluded from the BO tuning process. The detailed BO configuration, parameter ranges, number of training and test profiles, and numerical performance are reported in § 3.

Numerical Analysis

This section evaluates the proposed coordinated control framework for AI data center power smoothing and mitigation of grid-sensitive oscillatory load components. The controller parameters are tuned using BO over multiple stochastic AI workload realizations, and the resulting policies are subsequently evaluated on unseen load profiles. The numerical analysis focuses on grid-side power shaping, BO convergence, frequency response, and attenuation of the dominant oscillatory load component.

Simulation Setup and Parameter Settings

Aggregate Power Grid and AI Data Center Load Model

To evaluate the impact of highly variable AI workloads on grid frequency stability, the data center is connected to an aggregate power system represented by linearized swing and governor dynamics. The grid parameters are selected such that the resulting system contains a poorly damped natural mode at fnat=0.5f_{\mathrm{nat}} = 0.5 Hz.

The AI data center demand profile, Pdc(t)P_{\mathrm{dc}}(t), is generated using an Ornstein–Uhlenbeck (OU) stochastic process combined with compound Poisson burst events [9]. The OU component represents continuously varying computational utilization, whereas the Poisson component introduces intermittent workload surges. In addition, a periodic component with an amplitude of 20 MW at 0.5 Hz is superimposed on the stochastic demand to explicitly excite the vulnerable grid mode. Each simulation covers 300 s with a sampling interval of Δt=0.25\Delta t = 0.25 s.

HESS Configuration and Parameters

The Hybrid Energy Storage System (HESS) consists of a BESS with a power limit of 40 MW and an operational energy range of 4–36 MWh, together with a supercapacitor (SC) rated at 20 MW with an operational energy range of 0.03–0.27 MWh. Their initial stored energies are 20 MWh and 0.15 MWh, respectively. The BESS has a converter response time constant of τbat=0.08\tau_{\mathrm{bat}} = 0.08 s and the SC has a faster response of τsc=0.05\tau_{\mathrm{sc}} = 0.05 s.

The storage dynamics account for charging/discharging efficiencies and enforce both rated-power limits and instantaneous power limits imposed by the available stored energy. Therefore, neither storage device can discharge below its minimum energy level or charge beyond its maximum energy level. The principal system parameters and objective coefficients are summarized in Table I [8].

CategoryParameterValue
Grid & LoadBase Load (PbaseP_{\mathrm{base}})25.0 MW
Grid & LoadGrid Power Envelope[120.0, 130.0] MW
Grid & LoadRamp Rate Limit (Rg,maxR_{\mathrm{g,max}})1.0 MW/s
Grid & LoadGrid Natural Freq. (fnatf_{\mathrm{nat}})0.5 Hz
BESSPower Limit (P‾bat\overline{P}_{\mathrm{bat}})40.0 MW
BESSEnergy Limits (E‾bat,E‾bat\underline{E}_{\mathrm{bat}}, \overline{E}_{\mathrm{bat}})4.0, 36.0 MWh
BESSEfficiency (ηbat\eta_{\mathrm{bat}}), Response (τbat\tau_{\mathrm{bat}})0.97, 0.08 s
BESSDegradation Cost (cbatdegc_{\mathrm{bat}}^{\mathrm{deg}})$20.0/MWh
BESSRamping Penalty (cbatrampc_{\mathrm{bat}}^{\mathrm{ramp}})$2.0/(MW/s)2^2h
SCPower Limit (P‾sc\overline{P}_{\mathrm{sc}})20.0 MW
SCEnergy Limits (E‾sc,E‾sc\underline{E}_{\mathrm{sc}}, \overline{E}_{\mathrm{sc}})0.03, 0.27 MWh
SCEfficiency (ηsc\eta_{\mathrm{sc}}), Response (τsc\tau_{\mathrm{sc}})0.98, 0.05 s
SCDegradation Cost (cscdegc_{\mathrm{sc}}^{\mathrm{deg}})$5.0/MWh
SCRamping Penalty (cscrampc_{\mathrm{sc}}^{\mathrm{ramp}})$0.2/(MW/s)2^2h
Obj. WeightsEnv. Violation Penalty (cpc_{\mathrm{p}})$10.0/MW2^2h
Obj. WeightsRamp Violation Penalty (crc_{\mathrm{r}})$3.0/(MW/s)2^2h
Obj. WeightsModal Power Penalty (cmc_{\mathrm{m}})$100.0/MW2^2h

Table I. System Parameters and Cost Coefficients

3) Bayesian Optimization and Evaluation Protocol: The controller contains five tunable parameters,

θ=[τLPF,αsc,Kmode,Ksoc,bat,Ksoc,sc],\boldsymbol{\theta} = [\tau_{\mathrm{LPF}}, \alpha_{\mathrm{sc}}, K_{\mathrm{mode}}, K_{\mathrm{soc,bat}}, K_{\mathrm{soc,sc}}],

with corresponding search intervals

τLPF∈[0.5,20],αsc∈[0,1],\tau_{\mathrm{LPF}} \in[0.5, 20], \qquad\alpha_{\mathrm{sc}} \in[0, 1],
Kmode∈[0,10],Ksoc,bat∈[0,10],Ksoc,sc∈[0,100].K_{\mathrm{mode}} \in[0, 10], \qquad K_{\mathrm{soc,bat}} \in[0, 10], \qquad K_{\mathrm{soc,sc}} \in[0, 100].

Rather than optimizing the controller for a single stochastic realization, the BO objective is defined as the mean control cost over 10 fixed training workload profiles,

Jtrain(θ)=110∑s=110J(θ,Pdc(s)).J_{\mathrm{train}}(\boldsymbol{\theta}) = \frac{1}{10}\sum_{s=1}^{10} J\left(\boldsymbol{\theta}, P_{\mathrm{dc}}^{(s)}\right).

This formulation reduces overfitting to a particular workload realization. A Gaussian-process surrogate with normalized inputs and standardized outputs is employed together with the logarithmic Expected Improvement acquisition function. The optimization begins with 20 randomly sampled initial designs, followed by 50 sequential BO evaluations, resulting in 70 objective-function evaluations per case.

After optimization, the selected parameters are evaluated on 20 independent workload realizations that are not used during training. The time-domain and frequency-domain results presented below correspond to the first unseen test realization.

To assess the contribution of the individual objective components, three cases are considered:

  • Case 1 (Complete Model): power-envelope, grid-ramp, and modal penalties are all included in the optimization objective.

  • Case 2 (Mode Suppression Only): the explicit power-envelope and grid-ramp penalties are disabled, while the 0.5 Hz modal penalty is retained.

  • Case 3 (No Modal Penalty): the power-envelope and grid-ramp penalties are retained, while the explicit modal penalty is disabled.

It should be noted that the controller structure itself retains the low-pass power-sharing mechanism in all three cases. Consequently, Cases 2 and 3 still exhibit a degree of general power smoothing even when the corresponding objective terms are disabled.

Performance of the Proposed Coordinated Control

Figure 2 shows the raw AI data center demand, the resulting grid-side power, and the BESS and SC power trajectories for the three cases. The considered unseen test realization contains pronounced stochastic variations superimposed on the 0.5 Hz periodic component, with the raw demand varying approximately between 75 MW and 160 MW.

Figure 2. Power-smoothing performance on an unseen AI workload realization. Raw data center demand Pdc(t)P_{dc}(t), resulting grid-side power Pg(t)P_g(t), BESS power Pbat(t)P_{bat}(t), and SC power Psc(t)P_{sc}(t) are shown for Cases 1–3. Dashed horizontal lines indicate the 120–130 MW desired grid-power envelope.

For Case 1, the HESS substantially reshapes the raw demand and maintains PgP_g close to the desired 120–130 MW operating region for most of the simulation. Only short-duration excursions occur during the strongest workload bursts. Case 3 exhibits similarly strong low-frequency power regulation because the power-envelope and ramp penalties remain active. Case 2 also keeps the grid-side power centered near the desired operating region, despite the removal of the explicit envelope penalty. This behavior results from the underlying controller structure, which includes the low-pass reference term (pLPF−125p_{\mathrm{LPF}} - 125) and therefore inherently regulates the grid-side power around 125 MW. However, Case 2 exhibits more noticeable transient excursions than Case 1.

The complementary operation of the two storage technologies is also visible in Figure 2. The SC responds predominantly to the fast periodic component, repeatedly charging and discharging at high frequency, whereas the BESS provides the slower energy-balancing component associated with sustained deviations of the workload from the grid-side reference. This separation reduces the exposure of the BESS to the fastest power fluctuations while allowing the SC to absorb the high-frequency component.

Figure 3 presents the BO convergence behavior for the three cases. The first 15 evaluations correspond to the initial randomly sampled controller designs, after which the Gaussian-process surrogate and acquisition function are used to select new candidates. The individual candidate costs exhibit substantial variability, reflecting the nonlinear interaction between the five controller parameters. In contrast, the best-so-far objective decreases monotonically as progressively improved controller settings are discovered.

Figure 3. Bayesian Optimization convergence for Cases 1–3. The light curve shows the mean training cost of each evaluated controller, while the solid curve shows the best-so-far cost. The vertical dashed line marks the end of the 20-sample initial design and the beginning of the sequential BO stage.

For Case 1, the best mean training cost decreases rapidly during the initial design stage from above 200 to approximately 100, followed by a smaller additional improvement during the BO stage to approximately 90. Case 2 converges from an initial cost exceeding 350 to approximately 50, while Case 3 reaches a similar final value after a pronounced improvement shortly after the BO stage begins. These results indicate that the majority of the improvement is obtained

within a relatively small number of evaluations, while the later BO iterations primarily refine the solution.

The absolute objective values of the three cases should not be directly interpreted as a ranking of overall control quality because different penalty components are active in each case. The convergence curves are instead used to demonstrate that BO successfully identifies progressively improved parameter sets within each respective objective formulation.

Grid Frequency Response and Oscillation Mitigation

The resulting grid-frequency trajectories are shown in Figure 4. Without HESS compensation, the periodic 0.5 Hz component of the AI data center demand continuously excites the poorly damped grid mode, producing sustained oscillations of approximately 59.85–60.15 Hz after the initial transient. All three optimized controllers significantly reduce these oscillations. Cases 1 and 2 maintain the grid frequency very close to the nominal 60 Hz value, with only short transient deviations associated with stochastic workload bursts. Case 3 also provides a substantial reduction relative to the raw demand, but more visible residual oscillations remain during several intervals. The result demonstrates that generic power smoothing alone already provides substantial frequency support, while explicit consideration of the vulnerable mode further improves attenuation of the resonant component.

Grid-frequency response for the raw AI data center demand and the HESS-filtered grid-side power obtained under Cases 1–3 on the unseen test realization

Figure 4. Grid-frequency response for the raw AI data center demand and for the HESS-filtered grid-side power obtained under Cases 1–3 on the unseen test realization.

The frequency-domain behavior is examined further in Figure 5, which shows the five dominant non-DC spectral components of the raw and controlled power signals. The raw data center demand contains a dominant 20 MW component at precisely 0.5 Hz, corresponding to the imposed periodic computational-load oscillation. This component is considerably larger than the remaining stochastic spectral components and therefore dominates the excitation of the grid mode.

Figure 5. FFT comparison of the five dominant non-DC frequency components of the raw data center demand and the resulting grid-side power for Cases 1–3. The shaded region highlights the vulnerable grid-mode neighborhood around 0.5 Hz.

The HESS strongly attenuates this component in all three cases. In Case 1 and Case 2, the 0.5 Hz component no longer appears among the five dominant spectral components, indicating particularly strong suppression of the targeted mode. The remaining dominant components are primarily located at lower frequencies and remain below approximately 1 MW. Case 3 also achieves substantial attenuation because the low-pass power-sharing controller removes a large fraction of the fast load variation. However, a residual component remains visible near 0.5 Hz.

These results refine the interpretation of the ablation study. Explicit modal awareness is not solely responsible for all oscillation attenuation, since the HESS power-smoothing structure itself already suppresses a significant portion of the resonant variation. Nevertheless, including the modal penalty further reduces the spectral content around the vulnerable grid mode. Case 1 therefore provides the most complete formulation by jointly considering grid-power regulation, ramp-rate performance, storage operation, and mode-sensitive oscillation mitigation.

Conclusion

This study demonstrates that a coordinated HESS effectively mitigates extreme AI data center power volatility. By separating high-frequency transients from low-frequency demand shifts, the control policy leverages supercapacitors and batteries to optimize performance and minimize storage degradation.

Crucially, numerical analysis proves explicit targeted mode suppression is more vital for grid frequency stability than generic load smoothing. The optimized framework strictly regulates raw data center demand primarily within the 120–130 MW operating region, allowing only brief deviations during severe workload bursts. Targeting the vulnerable 0.5 Hz grid mode stabilizes sustained deviations and maintains frequency near the nominal 60 Hz. Furthermore, offline BO effectively navigates stochastic AI workload uncertainty, showing strong convergence by reducing mean training costs. This offers a robust, computationally efficient, and forecasting-independent solution for grid-compliant operation without real-time prediction.

References

  1. [1]A. Shehabi, S. J. Smith, A. Hubbard, A. Newkirk, N. Lei, M. A. B. Siddik, B. Holecek, J. Koomey, E. Masanet, and D. Sartor, “2024 united states data center energy usage report,” Lawrence Berkeley National Laboratory, Berkeley, CA, USA, Tech. Rep. LBNL-200601637, Dec. 2024.
  2. [2]G. Valverde, G. Pierrou, J. Krajacic, and G. Hug, “Forced oscillations in power systems induced by ai workloads,” 2025.
  3. [3]M.-S. Ko and H. Zhu, “Wide-area power system oscillations from large-scale ai workloads,” IEEE Transactions on Power Systems, 2026.arxiv.org/abs/2508.16457
  4. [8]X. Chen, “Grid-mode-aware model predictive control of hybrid energy storage systems for a data center power smoothing,” arXiv preprint arXiv:2609.04398, 2026.arxiv.org/abs/2609.04398
  5. [9]S. Lu, C. Xiao, and Y. Weng, “Dynamic load model for data centers with pattern-consistent calibration,” arXiv preprint arXiv:2602.07859, 2026.arxiv.org/abs/2602.07859

Paper details

vlnc:W_zddznzxm
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 23

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents