Abstract

The vast expansion of data center campuses has created concentrated and highly dynamic electrical loads that challenge traditional transmission system planning and protection studies. This paper presents a novel approach to model the load of data centers and validate their dynamic performance in a composite-load-model (CMLD) framework.

The data center has three major modules: a static element representing the static devices and busway impedance; an electronic-based element constituting power converters such as rectifiers and inverters, associated power supplies, and a motor-based element encompassing the induction nature of chillers, fans and pumps. Each module represented equivalent algebraic and differential equations depicting the dynamic behavior of the load.

A data-driven approach was used to predict the CMLD parameters of a data center under a fault condition. The procedure included curating real data centers events, estimating the CMLD parameters ranges, and automating the process of generating dynamic files through Latin Hypercube Sampling. We trained the simulation data using four machine learning groups of models: ensemble-tree models, kernel-based models, Gaussian processes models, and linear models. The trained models were then used to make inverse CMLD predictions on real and unseen data center events, evaluating the predictions, and selecting the most possible set of CMLD parameters for that event.

Compared with a base reference model, the procedure showed significant improvement in reducing the error between the real event and the trained models’ predictions. The methodology was validated on other data centers events under varying disturbance conditions.

Meeting Name

2025 Grid of the Future Symposium, Denver Colorado

Department(s)

Electrical and Computer Engineering

Keywords and Phrases

Composite load modeling, data centers, Gaussian processes, kernel machines, machine learning

Document Type

Article - Conference proceedings

Document Version

Citation

File Type

text

Language(s)

English

Publication Date

December 2025

Download

Full Text Link

Share

 
COinS