An Adaptive Multi-Agent Framework for Energy Optimization in Green Data Centers

Data centers account for approximately 1-2% of global electricity consumption, with projections reaching 8% by 2030. Traditional management systems rely on reactive control mechanisms that cannot adapt to dynamic conditions or renewable energy fluctuations. This paper presents an intelligent framework using multi-agent deep reinforcement learning to optimize energy use while maintaining service quality. The system coordinates three specialized agents controlling cooling infrastructure, workload scheduling, and renewable energy storage. Graph neural networks model thermal dynamics, replacing expensive computational fluid dynamics simulations. Unlike existing solutions that optimize subsystems independently, this integrated approach discovers synergistic strategies across thermal management, computation, and energy storage. We incorporate carbon-awareness directly into optimization objectives, enabling intelligent workload shifting to periods of high renewable availability. Safety constraints through Lagrangian-based reinforcement learning guarantee thermal limits and service agreements are maintained. This framework provides a blueprint for achieving carbon-neutral data center operations while preserving performance and reliability.