Research map / Multi-agent systems and games / Cooperative multi-agent RL
Carbon Footprint Reduction for Sustainable Data Centers in Real-Time
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DC-CFR, a multi-agent PPO/IPPO-based reinforcement learning framework that jointly controls data center load shifting, HVAC cooling setpoints, and battery storage in real time to cut carbon emissions, energy use, and cost versus the ASHRAE controller.
Problem
Reducing data center carbon footprint in real time by jointly optimizing IT load shifting, cooling, and battery storage under varying weather and grid carbon intensity, which static day-ahead pipelines handle poorly.
Approach
Decomposes data center control into three coupled MDPs (load shifting, HVAC energy, battery operation), each handled by a DRL agent that shares state variables and uses a collaborative reward mixing all agents' contributions, trained with PPO in a Sinergym/EnergyPlus simulator.
Key results
- DC-CFR reduced carbon footprint by 14.46% on average versus ASHRAE across locations (AZ 14.36%, NY 15.08%, WA 13.96%).
- DC-CFR reduced energy consumption by 14.35% on average and energy cost by 13.69% versus ASHRAE.
- Combined agents outperform individual strategies; battery agent alone gives under 0.5% carbon reduction, while load shifting plus HVAC optimization gives about 13-14%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine load shifting, HVAC energy optimization, and battery operation in DC-CFR rather than using LS+EO (ablation) | DC-CFR (IPPO) · Arizona, 1-year simulation | carbon footprint reduction vs ASHRAE (%): 13.26 ± 0.07 -> 14.36 ± 0.09 | Helped |
| Combine load shifting, HVAC energy optimization, and battery operation in DC-CFR rather than using LS+EO (ablation) | DC-CFR (IPPO) · New York, 1-year simulation | carbon footprint reduction vs ASHRAE (%): 14.39 ± 0.08 -> 15.08 ± 0.11 | Helped |
| Combine load shifting, HVAC energy optimization, and battery operation in DC-CFR rather than using LS+EO (ablation) | DC-CFR (IPPO) · Washington, 1-year simulation | carbon footprint reduction vs ASHRAE (%): 13.62 ± 0.08 -> 13.96 ± 0.06 | Helped |
| Combine load shifting, HVAC energy optimization, and battery operation in DC-CFR rather than using LS+EO (ablation) | DC-CFR (IPPO) · Arizona, 1-year simulation | energy consumption reduction vs ASHRAE (%): 14.28 ± 0.07 -> 14.54 ± 0.33 | Helped |
| Combine load shifting, HVAC energy optimization, and battery operation in DC-CFR rather than using LS+EO (ablation) | DC-CFR (IPPO) · New York, 1-year simulation | energy consumption reduction vs ASHRAE (%): 14.35 ± 0.08 -> 14.62 ± 0.07 | Helped |
| Combine load shifting, HVAC energy optimization, and battery operation in DC-CFR rather than using LS+EO (ablation) | DC-CFR (IPPO) · Washington, 1-year simulation | energy consumption reduction vs ASHRAE (%): 13.78 ± 0.06 -> 13.85 ± 0.07 | Helped |
| Combine load shifting, HVAC energy optimization, and battery operation in DC-CFR rather than using LS+EO (ablation) | DC-CFR (IPPO) · Arizona, 1-year simulation | energy cost reduction vs ASHRAE (%): 13.81 ± 0.11 -> 14.07 ± 0.17 | Helped |
| Combine load shifting, HVAC energy optimization, and battery operation in DC-CFR rather than using LS+EO (ablation) | DC-CFR (IPPO) · New York, 1-year simulation | energy cost reduction vs ASHRAE (%): 13.61 ± 0.09 -> 14.16 ± 0.11 | Helped |
| Combine load shifting, HVAC energy optimization, and battery operation in DC-CFR rather than using LS+EO (ablation) | DC-CFR (IPPO) · Washington, 1-year simulation | energy cost reduction vs ASHRAE (%): 12.81 ± 0.05 -> 12.85 ± 0.05 | Helped |
| Add battery operation to load shifting and HVAC energy optimization (ablation) | DC-CFR (IPPO) · All three locations, 1-year simulation | energy consumption reduction vs ASHRAE (%): 0.00 ± 0.00 | No clear effect |
Surprising findings
- The battery agent alone yields only 0.25-0.46% carbon reduction, but its contribution is meaningful when combined with other agents' policies.
Limitations
- Evaluated only in simulation (EnergyPlus/Sinergym), not on a real data center.
- Trained on New York data and tested on other locations; generalization is limited to three locations.
- Embodied carbon of servers and batteries is ignored.
Experimental setup
Datasets: Alibaba cluster trace IT load data New York weather and grid carbon intensity data Arizona and Washington weather and carbon intensity data
Benchmarks: EnergyPlus two-zone data center simulation via Sinergym (Arizona, New York, Washington, one year)
Metrics: percentage reduction in carbon footprint percentage reduction in energy consumption percentage reduction in energy cost
Baselines: ASHRAE rule-based controller MADDPG
Builds on: MADDPG PPO Sinergym Carbon Explorer
Closest papers by meaning
- Hierarchical Multi-Agent Framework for Carbon-Efficient Liquid-Cooled Data Center Clusters (AAAI 2025 · similarity 0.92)
- SustainDC: Benchmarking for Sustainable Data Center Control (NeurIPS 2024 · similarity 0.92)
- Sustainability of Data Center Digital Twins with Reinforcement Learning (AAAI 2024 · similarity 0.91)
- DCcluster-Opt: Benchmarking Dynamic Multi-Objective Optimization for Geo-Distributed Data Center Workloads (NeurIPS 2025 · similarity 0.90)
- Data Center Cooling System Optimization Using Offline Reinforcement Learning (ICLR 2025 · similarity 0.88)
- LC-Opt: Benchmarking Reinforcement Learning and Agentic AI for End-to-End Liquid Cooling Optimization in Data Centers (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.