Advances in power consumption model for data centers: Analytical formulas vs. machine learning models

Cloud computing services are increasingly driving the development and deployment of artificial intelligence (AI) models, leveraging large computing infrastructures with ever-increasing energy consumption. The costs of building data centers with the required energy capacity can rival their ongoing en...

Ausführliche Beschreibung

Gespeichert in:

Bibliographische Detailangaben
Veröffentlicht in:	Future generation computer systems Jg. 176; S. 108141
Hauptverfasser:	Racedo, Sebastian, Jaumard, Brigitte, Glatard, Tristan, Delgado, Oscar, Masoudi, Meysam
Format:	Journal Article
Sprache:	Englisch
Veröffentlicht:	Elsevier B.V 01.03.2026
Schlagworte:	Analytical formula Comparison Data center Machine learning models Power consumption estimation Analytical formula Power consumption estimation Data center Machine learning models Comparison
ISSN:	0167-739X
Online-Zugang:	Volltext
Tags:	Tag hinzufügen Keine Tags, Fügen Sie den ersten Tag hinzu!

Beschreibung
Zusammenfassung:	Cloud computing services are increasingly driving the development and deployment of artificial intelligence (AI) models, leveraging large computing infrastructures with ever-increasing energy consumption. The costs of building data centers with the required energy capacity can rival their ongoing energy expenditures, creating strong economic incentives to operate these facilities at or near full capacity. However, achieving this efficiency remains challenging due to uncertainties in equipment power ratings and significant fluctuations in power consumption depending on server workloads. Accurate and fast power consumption modeling is therefore critical essential to optimize these infrastructures, yet existing analytical methods often need to be refined to adapt to high-end servers. In this work, we propose and rigorously validate a data-driven closed-form formula for power estimation. While the mathematical form relies on polynomial regression, the novelty lies in its demonstrated stability across multiple machines, its strong accuracy, and its ability to generalize with limited calibration data. We evaluate its performance against several machine learning models, including XGBoost (tree- and linear-based), Ridge regression, and Multi-layer Perceptrons to estimate power consumption. All of the methods (proposed formula and ml models) are evaluated on data collected from a dedicated testbed comprising multiple physical machines, ensuring broader generalization beyond a single machine. The results highlight the effectiveness of our improved closed-form approach, as well as its potential advantages over machine learning-based solutions for power estimation.
ISSN:	0167-739X
DOI:	10.1016/j.future.2025.108141