Leveraging Deep Reinforcement Learning within Optimal Renewable Energy Strategies for Sustainable AI Data Centers
Tianqi Xiao, Fengqi You · Environmental Science & Technology · 2025
AI computing's rapid expansion is steeply increasing data center electricity use, intensifying sustainability concerns. We develop the first framework that couples deep reinforcement learning (DRL) control with cost-effective optimization to boost efficiency and enable economically viable renewable integration in AI data centers. Using seven public, real-world AI workloads and up-to-date open-source grid and renewable-cost data sets, we assess energy, water, and carbon performance at ten globally representative sites, benchmarking against an ASHRAE standard-aligned baseline controller. DRL attains near-optimal free-cooling operation, delivering over 6% energy reduction and over 8% water savings. Sustaining higher server utilization could further cut auxiliary cooling by up to 60% per unit of server energy when wet-bulb temperatures exceed the free-cooling thresholds. We also evaluate price- and carbon-oriented demand response potentials combined with battery storage. The integrated strategy yields concurrent cost and emission reductions, lowering the total cost of a 50% emission cut by 9-28% and placing abatement costs at $107-$500 per ton for on-site renewable adoption across selected locations. These results show that intelligent controls, paired with renewable strategies, can deliver scalable, cost-effective decarbonization of AI infrastructure consistent with global efficiency and net-zero goals.