Scholarlink

With data centres using 1–2% of the world's energy, efficiency in cooling is the key to sustainability. Legacy cooling technology based on static threshold values cannot adapt to dynamic workloads and climatic conditions, which causes inefficiency. In this paper, we present deep Q-learning with Experience Replay and Bellman optimization for optimal energy spending with guaranteed server safety temperatures. In a simulated data centre environment with realistic parameters, our AI model learns to adaptively adjust cooling setpoints, saving 21% energy spending compared to traditional practices. Contributions include a hybrid reward mechanism, which will be discussed in detail in the methodology, and an extensible real-time monitoring system. Variable climatic condition experiments prove the model's learnability and adaptability, resulting in a scalable, AI-based solution that realizes maximum energy efficiency over traditional techniques such as PUE. The system meets UN SDG 7 and 13 and presents actionable data centre operator and policymaker recommendations.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Reinforcement Learning for Dynamic Energy Optimization in Data Centers

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shrivatsa Deshpande,
  • Parth Gala,
  • Dipti Pawade,
  • Avani Bhattacharjee,
  • Manan Kabra

摘要

With data centres using 1–2% of the world's energy, efficiency in cooling is the key to sustainability. Legacy cooling technology based on static threshold values cannot adapt to dynamic workloads and climatic conditions, which causes inefficiency. In this paper, we present deep Q-learning with Experience Replay and Bellman optimization for optimal energy spending with guaranteed server safety temperatures. In a simulated data centre environment with realistic parameters, our AI model learns to adaptively adjust cooling setpoints, saving 21% energy spending compared to traditional practices. Contributions include a hybrid reward mechanism, which will be discussed in detail in the methodology, and an extensible real-time monitoring system. Variable climatic condition experiments prove the model's learnability and adaptability, resulting in a scalable, AI-based solution that realizes maximum energy efficiency over traditional techniques such as PUE. The system meets UN SDG 7 and 13 and presents actionable data centre operator and policymaker recommendations.