Data Center Thermal Optimization Using Graph Neural Networks

Abstract

Data center cooling accounts for a significant fraction of facility energy consumption, motivating the development of intelligent control strategies that improve efficiency while maintaining thermal safety. This study presents a graph-based digital twin for data center thermal optimization that enables safe online reinforcement learning for cooling control. A Graph Attention Network v2 (GATv2) models node-level thermal dynamics by capturing spatial interactions between neighboring server nodes and serving as the simulation environment for controller training. The proposed model outperformed LSTM, GRU, GCN, and GAT baselines, achieving the highest outlet temperature prediction accuracy. Reinforcement learning policies trained within the digital twin maintained zero thermal limit violations while reducing cooling demand relative to the historical controller. Among the evaluated algorithms, Soft Actor-Critic achieved the best overall balance between fan energy savings and cooling valve reduction, reducing both fan energy consumption and cooling valve opening by approximately 20%.