Thermal prediction for Air-cooled data center using data Driven-based model. (25th November 2022)

• Record Type: Journal Article
• Authors: Lin, Jianpeng; Lin, Weiwei; Lin, Wenjun; Wang, Jiangtao; Jiang, Hongliang
• Is Part Of: Applied thermal engineering. Volume 217(2022)
• Publication Date: 2022-11-25
• Subjects: Thermal modeling -- Air-cooled data center -- Temperature prediction -- Machine learning; Heat engineering -- Periodicals; Heating -- Equipment and supplies -- Periodicals; Periodicals; 621.40205
• DOI: 10.1016/j.applthermaleng.2022.119207
• Languages: English
• ISSNs: 1359-4311

Abstract

Highlights: This work comprehensively compares the temperature prediction performance of six data-driven-based thermal models in steady-state and transient-state scenarios. Extensive numerical experiments are conducted to explore how sample size, physical layout reconfiguration, and airflow patterns affect thermal model performance. Four cooling failure transient scenarios are designed to verify the ability of the thermal model to capture the thermal evolution. XGBoost and LightGBM perform better than other thermal models with RMSE less than 1.0 °C and time overhead less than 10 -3 s. Abstract: The optimal cooling control of data centers (DCs) relies on the thermal model to simulate and accurately evaluate the temperature distribution of the computer room. Data-driven thermal modeling methods have been widely used in the thermal management of DCs. However, few existing works have comprehensively compared and analyzed the predictive performance and adaptability of various data-driven thermal models. In addition, most of the proposed thermal models are for temperature prediction in steady-state scenarios of DCs, ignoring the study of thermal changes in transient scenarios. Therefore, this work builds a CFD model of a typical data center as a verification platform to compare the temperature prediction performance of six machine learning (ML)-based thermal models (SVR, GPR, XGBoost, LightGBM, ANN, LSTM) in steady-state and transient-state scenarios. For steady-state scenarios, we also explore the impact of sample size, physical layout reconfiguration, and airflow pattern on performance to evaluate the adaptability of the thermal model. For transient scenarios, we design four cooling failure scenarios to verify the ability of the proposed thermal model to capture the dynamic thermal changes of the computer room. Numerous numerical simulation results show that the XGBoost-based and LightGBM-based thermal models outperform other ML models in both steady-state and transient-state scenarios, with prediction error RMSE is less than 1.0 ℃, and the time overhead is less than 10 -3 s. In addition, both thermal models are robust to physical layout reconfiguration and poor flow patterns.