A research team has developed a deep reinforcement learning framework designed to coordinate mixed artificial intelligence workloads within data centers powered by combined on-site photovoltaic generation and external power grids. The scheduling method addresses persistent operational challenges created by temporal mismatches between intermittent solar power generation and the electricity demands of computing hardware.
Artificial intelligence infrastructure faces rising electrical energy consumption, which increases the necessity of integrating local renewable power sources. Managing these computing facilities requires coordinating diverse workloads with competing timing requirements, specifically balancing latency-sensitive inference requests alongside computationally heavy training tasks that possess higher temporal scheduling flexibility.
To navigate these conflicting demands across heterogeneous multi-graphics processing unit server environments, the researchers implemented a scheduler using a dueling double deep Q-network architecture. The model dynamically allocates incoming computational tasks by continuously assessing short-term solar generation forecasts, specific workload profiles, and real-time graphics processing unit resource availability states.
Under this scheduling strategy, inference tasks are executed promptly without unnecessary processing delays, while deferrable model training jobs are scheduled strategically to align execution windows with peak periods of available solar photovoltaic power.
Experimental evaluations demonstrated that the dueling double deep Q-network approach improved the green energy consumption ratio while maintaining low service-level agreement violation rates across multiple workload configurations when compared against existing scheduling approaches.
Subject of Research: Green-energy-aware scheduling of mixed AI workloads in data centers using deep reinforcement learning
Article Title: Deep reinforcement learning for green-energy-aware scheduling of mixed AI workloads in data centers
Article References: Ma, H., Jiang, H., Pei, H., & Cheng, L. (2026). Deep reinforcement learning for green-energy-aware scheduling of mixed AI workloads in data centers. Journal of Cloud Computing. https://doi.org/10.1186/s13677-026-01024-y
Image Credits: AI Generated
DOI: 10.1186/s13677-026-01024-y
Keywords: data center scheduling, deep reinforcement learning, photovoltaic power, dueling double deep Q-network, AI workloads, mixed workloads, inference tasks, model training, green computing, GPU servers, renewable energy integration, service-level agreements, grid electricity, energy-aware computing, heterogeneous servers, short-term solar forecasting
News Source: Denise Maddox. (November 21, 2026). Deep Reinforcement Learning Optimizes Green Energy Scheduling for Mixed AI Data Workloads. Scienmag.



