Reproduce “Data Center Cooling System Optimization Using Offline Reinforcement Learning” on OpenResearch

About this paper

Xianyuan Zhan, Xiangyu Zhu, Peng Cheng, Xiao Hu, Ziteng He, Hanfei Geng, Jichao Leng, Huiwen Zheng and 5 more

Researchers developed a physics-informed offline reinforcement learning framework for data center cooling systems. This framework integrates a T-symmetry enforced graph neural network to model thermal dynamics, resulting in a 14% to 21% reduction in air-side cooling load factor in a commercial data center while ensuring thermal safety and operational stability.

Key insights

• Integrating physical principles, specifically time-reversal symmetry, into GNN-based dynamics models significantly boosts data efficiency and generalization in offline RL, even with limited industrial data.
• Learning control policies within a robust, physics-informed latent space enables safe and effective optimization for real-world safety-critical systems, circumventing the need for high-fidelity simulators or risky online interaction.
• Careful design of a safety-aware reward function combined with policy regularization is essential for maintaining strict operational constraints while optimizing for energy efficiency in complex industrial environments.

Summary by alphaXiv.