Leveraging Deep Reinforcement Learning within Optimal Renewable Energy Strategies for Sustainable AI Data Centers

Area of Science:

• Artificial Intelligence
• Sustainable Computing
• Data Center Operations

Background:

• AI computing's growth drives significant data center energy consumption, raising sustainability issues.
• Existing data center control methods lack efficiency and cost-effective renewable integration.
• Decarbonizing AI infrastructure is crucial for meeting global net-zero targets.

Purpose of the Study:

• To develop and evaluate a novel framework integrating deep reinforcement learning (DRL) with cost-effective optimization for AI data centers.
• To enhance energy efficiency, reduce water consumption, and enable economically viable renewable energy integration.
• To assess the potential for cost and emission reductions through intelligent controls and renewable strategies.

Main Methods:

• Developed a framework coupling DRL control with cost-effective optimization.
• Utilized seven real-world AI workloads and open-source grid/renewable cost data.
• Assessed energy, water, and carbon performance across ten global sites.
• Benchmarked against an ASHRAE standard-aligned baseline controller.
• Evaluated demand response and battery storage integration.

Main Results:

• DRL achieved near-optimal free-cooling, yielding over 6% energy and 8% water savings.
• Higher server utilization could reduce auxiliary cooling by up to 60% under specific conditions.
• Integrated strategies reduced the total cost of a 50% emission cut by 9-28%.
• Abatement costs for on-site renewables ranged from $107-$500 per ton across locations.

Conclusions:

• Intelligent DRL controls coupled with renewable strategies offer scalable and cost-effective decarbonization for AI infrastructure.
• The developed framework aligns with global efficiency and net-zero goals for AI data centers.
• This approach demonstrates a viable pathway to reduce the environmental impact of AI computing.