Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dynamic Load Model for Data Centers with Pattern-Consistent Calibration

Siyu Lu, Chenhan Xiao, Yang Weng

TL;DR

This work addresses the challenge of modeling large electronic loads (LELs) like data centers for grid stability by combining a physics-based, parameterized backbone with pattern-consistent calibration. It introduces Temporal Contrastive Learning (TCL) to align facility data with simulations on temporal and statistical patterns, rather than forcing exact trajectory matches, enabling privacy-preserving, facility-level calibration of IT workload, cooling, and auxiliary subsystems alongside protection settings. The approach is validated with real-world traces from MIT Supercloud, ASU Sol, Blue Waters, and ASHRAE data, integrated into the ANDES platform, and tested on IEEE 39-bus, NPCC 140-bus, and WECC 179-bus systems, revealing that interactions among heterogeneous LELs can cause compound disconnections and delayed stabilization not captured by uncalibrated models. The results show improved shape similarity to real loads, robustness to parameter initialization, and scalable accuracy across large grids, highlighting the method’s practical impact for planning, protection assessment, and grid resilience in the face of rising LEL penetration.

Abstract

The rapid growth of data centers has made large electronic load (LEL) modeling increasingly important for power system analysis. Such loads are characterized by fast workload-driven variability and protection-driven disconnection and reconnection behavior that are not captured by conventional load models. Existing data center load modeling includes physics-based approaches, which provide interpretable structure for grid simulation, and data-driven approaches, which capture empirical workload variability from data. However, physics-based models are typically uncalibrated to facility-level operation, while trajectory alignment in data-driven methods often leads to overfitting and unrealistic dynamic behavior. To resolve these limitations, we design the framework to leverage both physics-based structure and data-driven adaptability. The physics-based structure is parameterized to enable data-driven pattern-consistent calibration from real operational data, supporting facility-level grid planning. We further show that trajectory-level alignment is limited for inherently stochastic data center loads. Therefore, we design the calibration to align temporal and statistical patterns using temporal contrastive learning (TCL). This calibration is performed locally at the facility, and only calibrated parameters are shared with utilities, preserving data privacy. The proposed load model is calibrated by real-world operational load data from the MIT Supercloud, ASU Sol, Blue Waters, and ASHRAE datasets. Then it is integrated into the ANDES platform and evaluated on the IEEE 39-bus, NPCC 140-bus, and WECC 179-bus systems. We find that interactions among LELs can fundamentally alter post-disturbance recovery behavior, producing compound disconnection-reconnection dynamics and delayed stabilization that are not captured by uncalibrated load models.

Dynamic Load Model for Data Centers with Pattern-Consistent Calibration

TL;DR

This work addresses the challenge of modeling large electronic loads (LELs) like data centers for grid stability by combining a physics-based, parameterized backbone with pattern-consistent calibration. It introduces Temporal Contrastive Learning (TCL) to align facility data with simulations on temporal and statistical patterns, rather than forcing exact trajectory matches, enabling privacy-preserving, facility-level calibration of IT workload, cooling, and auxiliary subsystems alongside protection settings. The approach is validated with real-world traces from MIT Supercloud, ASU Sol, Blue Waters, and ASHRAE data, integrated into the ANDES platform, and tested on IEEE 39-bus, NPCC 140-bus, and WECC 179-bus systems, revealing that interactions among heterogeneous LELs can cause compound disconnections and delayed stabilization not captured by uncalibrated models. The results show improved shape similarity to real loads, robustness to parameter initialization, and scalable accuracy across large grids, highlighting the method’s practical impact for planning, protection assessment, and grid resilience in the face of rising LEL penetration.

Abstract

The rapid growth of data centers has made large electronic load (LEL) modeling increasingly important for power system analysis. Such loads are characterized by fast workload-driven variability and protection-driven disconnection and reconnection behavior that are not captured by conventional load models. Existing data center load modeling includes physics-based approaches, which provide interpretable structure for grid simulation, and data-driven approaches, which capture empirical workload variability from data. However, physics-based models are typically uncalibrated to facility-level operation, while trajectory alignment in data-driven methods often leads to overfitting and unrealistic dynamic behavior. To resolve these limitations, we design the framework to leverage both physics-based structure and data-driven adaptability. The physics-based structure is parameterized to enable data-driven pattern-consistent calibration from real operational data, supporting facility-level grid planning. We further show that trajectory-level alignment is limited for inherently stochastic data center loads. Therefore, we design the calibration to align temporal and statistical patterns using temporal contrastive learning (TCL). This calibration is performed locally at the facility, and only calibrated parameters are shared with utilities, preserving data privacy. The proposed load model is calibrated by real-world operational load data from the MIT Supercloud, ASU Sol, Blue Waters, and ASHRAE datasets. Then it is integrated into the ANDES platform and evaluated on the IEEE 39-bus, NPCC 140-bus, and WECC 179-bus systems. We find that interactions among LELs can fundamentally alter post-disturbance recovery behavior, producing compound disconnection-reconnection dynamics and delayed stabilization that are not captured by uncalibrated load models.
Paper Structure (18 sections, 1 theorem, 13 equations, 13 figures, 1 table)

This paper contains 18 sections, 1 theorem, 13 equations, 13 figures, 1 table.

Table of Contents

  1. Introduction
  2. Parameterized Physics-Based Load Modeling of Data Centers
  3. Duty-Idle Workload Modeling
  4. Cooling and Auxiliary Load Modeling
  5. Grid-Interfacing Protection and Recovery Modeling
  6. Pattern-Consistent Calibration Framework
  7. Statistical Pattern-Consistent Calibration
  8. Rule-Based Specification of Protection Parameters
  9. Numerical Results
  10. Experiment Setup
  11. Validation of Pattern-Consistent Calibration
  12. Robustness to Parameter Initialization
  13. Effect and Sensitivity of the TCL Feature Module
  14. System-Level Dynamic Response with heterogeneous LEL
  15. Impact of Load Model Calibration
  16. ...and 3 more sections

Key Result

Lemma 1

Consider the core impulse-driven component of a data center workload modeling in Eq. (eq:eta): $I(t)=\sum_{k}\delta(t-t_k)$ over a finite time horizon $[0,T]$, where $\{t_k\}$ are the job-arrival times of a homogeneous Poisson process with rate $\lambda^{\ast}>0$. A single time-series sequence of th

Figures (13)

  • Figure 1: Pattern-consistent calibration framework for realistic LEL dynamic behavior in grid analysis with data privacy preserved.
  • Figure 2: Illustration of an OU process in Eq. (\ref{['eq:eta']}) and a Poisson process. The pure Poisson process is memoryless and too stochastic.
  • Figure 3: Pattern-consistent calibration of LEL modeling parameters.
  • Figure 4: Comparison between real CPU/GPU workload traces and simulated workload trajectories for three computing sites.
  • Figure 5: Empirical cooling load traces and simulated cooling trajectories.
  • ...and 8 more figures

Theorems & Definitions (2)

  • Lemma 1: Overfitting risk of trajectory alignment for data center workloads
  • proof