Increasing the energy efficiency of a data center based on machine learning

TLDR

Data centers consume large amounts of energy, and while past growth has shifted toward larger, more efficient facilities, further gains from more efficient operation are essential to curb future consumption and environmental impact, with machine learning offering promising control strategies. The study proposes ready‑to‑use machine learning methods with practical details to reduce power usage effectiveness (PUE) in data centers. These methods were implemented at Tencent Tianjin’s hyperscale data center, where they optimized cooling set points and corrected unreasonable operating conditions. The models—neural network, light gradient boosting machine, recurrent neural network, and random forest—accurately predicted PUE, achieving a minimum reduction of 0.005 that translates to about 1,500 MWh (0.24 %) saved annually, and the framework can be transferred to similarly sized centers.

Abstract

Abstract Energy efficiency of data centers (DCs) is of great concern due to their large amount of energy consumption and the foreseeable growth in the demand of digital services in the future. The past decade witnessed improvements of the energy efficiency of DCs from an extensive margin—a shift from small to large, more efficient DCs. Improvements from the intensive margin, that is, from more efficient operation, would be critical in limiting the energy consumption and environmental impact of DCs in the upcoming period. Machine learning algorithms have advantages in optimizing the operation of DCs to improve energy efficiency as they have shown the potential of discovering control strategies not found by traditional method, and producing working condition‐dependent control strategies. This study proposes ready‐to‐use machine learning methods with practical details to decrease the most commonly used energy efficiency metric—power usage effectiveness in DCs. We achieved an accurate prediction by properly selecting the features used in the proposed prediction models established by neural network, light gradient boosting machine, recurrent neural network, and random forests. The proposed approaches are implemented in one of the largest hyperscale DCs in China—Tencent Tianjin DC, to optimize the set points of controllable variables in the cooling system and to detect and adjust the unreasonable working conditions in the modular data centers. The lower bound of PUE reduction was 0.005 with the proposed approaches, leading to about 1500 MWh (0.24% of the total designed electricity consumption of this DC) of energy saved per year in this hyperscale DC. The proposed methods have the potential to be transferred to DCs of similar scale, and the framework of our work could serve as a guide for machine learning‐based optimization of environmental indicators in other complex product/service systems.