Deep reinforcement learning for green-energy-aware scheduling of mixed AI workloads in data centers
Year: 2026
Type: article
Abstract: The growing electricity demand of artificial intelligence (AI) data centers has made the effective use of on-site renewable energy increasingly important. However, temporal mismatches between renewable generation and computing demand pose significant challenges for green-energy-aware scheduling. Scheduling mixed AI workloads further complicates this task, as it requires accommodating latency-sensitive inference tasks while exploiting the temporal flexibility of training jobs. To address these challenges, we propose a deep reinforcement learning approach to workload scheduling in an AI data center powered by on-site photovoltaic (PV) generation and the external grid. Based on short-term PV forecasts, workload characteristics, and graphics processing unit (GPU) resource states, the scheduler assigns inference tasks and training jobs while considering the use of PV power and grid electricity. Specifically, in a scheduling process involving heterogeneous multi-GPU servers, inference tasks are completed without unnecessary delays and training jobs can be deferred in an effective way to make better use of available PV power. We present the detailed design of our approach and implement it using the dueling double deep Q-network. Experimental results show that our method can improve green energy consumption ratio and maintain low service-level agreement (SLA) violation rates across different workload configurations, compared to current approaches.
Authors: Hongzhe Ma, Hao Jiang, Haoran Pei, Long Cheng
Institutions: Wuhan University, North China Electric Power University
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,790
OpenAlex ID: W7237130644
