Optimal energy management for air cooled server fans using Deep Reinforcement Learning control method

Yogesh Fulpagare, Kuei Ru Huang, Ying Hao Liao, Chi Chuan Wang*

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

3 Scopus citations


The current study proposed the Deep Reinforcement Learning (DRL) AI control algorithm for energy-saving optimization in 1U air-cooled server mockup for different configurations and scenarios. The proposed algorithm can handle the server's complex thermal environment, including number of heat sources and their locations, type of heat sink, airflow, fan and bypass airflow. Based on the analysis and test results, it is found that controlling a single fan individually and the fans in different zones can offer the reduction of fan energy by 39–47% compared with the controlling all fans collectively. After the modification in observation characteristics, the fan usage rate was reduced by 45% compared with the single fan control model. Further, five different control models were developed and tested with experiments for different function variables and reward functions on three server configurations with different algorithmic settings for energy saving optimizations. The heat source of the three server configurations was close to the upper limit temperature and maintained stably by saving about 12% to 50% fan energy consumption. The algorithm can judge the system status and successfully provide real-time decision-making action thereby the energy consumption of the fan is significantly reduced.

Original languageEnglish
Article number112542
JournalEnergy and Buildings
StatePublished - 15 Dec 2022


  • Air cooling
  • Control
  • Data center
  • Deep reinforcement learning
  • Energy-saving


Dive into the research topics of 'Optimal energy management for air cooled server fans using Deep Reinforcement Learning control method'. Together they form a unique fingerprint.

Cite this