[Online]Beyond Airflow Magnitude: Distribution-Aware Machine Learning for Rack-Level Thermal Prediction in Data Centers

Beyond Airflow Magnitude: Distribution-Aware Machine Learning for Rack-Level Thermal Prediction in Data Centers
ID:122 Submission ID:10 View Protection:ATTENDEE Updated Time:2026-10-06 06:55:37 Hits:201 Online

Start Time:2026-10-12 16:00 (Asia/Ho_Chi_Minh)

Duration:15min

Session:[S5] Track 5: Emerging Trends of AI/ML » [S5-5] Track 5: Emerging Trends of AI/ML

Video No Permission Presentation File Attachment File

Tips: The file permissions under this presentation are only for participants. You have not logged in yet and cannot view it temporarily.

Abstract
Existing approaches for thermal management in data centers rely predominantly on computational fluid dynamics (CFD) simulations or facility-level machine learning models, which can be computationally demanding or may not explicitly represent rack-level airflow–thermal relationships. To address this challenge, this study proposes a predictive modeling framework for estimating rack-level thermal behavior from airflow and spatial characteristics using machine learning techniques. The analysis uses a reproducible synthetic dataset of 300 observations incorporating airflow velocity, airflow distribution index, and rack position. Linear Regression, Random Forest, and XGBoost models were evaluated using repeated 5-fold cross-validation with 10 repetitions, hyperparameter optimization, feature engineering, and an independent test set of 60 observations. Feature engineering introduced nonlinear and pairwise interaction terms among the airflow and spatial variables. The results show that Linear Regression with engineered features achieved the best independent-test performance, with R2=0.9392 and RMSE = 0.8538 °C. The engineered representation improved Linear Regression performance relative to the baseline, while providing a small training–test R2 gap of 0.0046. The results demonstrate the potential of combining airflow characteristics, spatial information, and engineered features for reproducible rack-level thermal prediction. Although the dataset is synthetic, the complete generation procedure was successfully reproduced, providing a basis for future validation using larger experimental, CFD derived, or operational datasets. The proposed framework therefore provides a data-driven foundation for future predictive thermal monitoring and distribution-aware cooling strategies in data center environments.
Keywords
machine learning, airflow distribution, data center cooling, rack-level thermal prediction, thermal management
Speaker
Cortes Antonio
Full Professor and Research Universidad de Panamá

Antonio CORTES CASTILLO is a computer engineer trained in the Latin American University of Science and Technology (ULACIT), Costa Rica, 1995. Obtained his PhD at the University of Alicante, 2020 and acquire his Master´s degree in Computer Science (Telematics) at the Technological Institute of Costa Rica, 2000. Besides, he obtained his bachelor's degree in computer engineering with an emphasis in Management Information Systems at the National University of Heredia, Costa Rica, 2002. He is now teaching and research at the University of Panama.

Submission Author
Antonio Cortes Universidad de Panamá
Comment submit
Verification code Change another
All comments

CONTACT US

Email: asiancomnet2026@usssociety.org

Website & IT Support: hi@aconf.org

Registration Submit Paper