Simulator-Based Reinforcement Learning for Data Center Cooling Optimization

Chi Zhou, Doris Gao, Lisa Rivalin, Andrew Grier, Gerson Arteaga Ramirez, John Fabian

Keywords: Reinforcement Learning, Physical Simulator, Data Center, Cooling Optimization

TL;DR: We present a simulator based reinforcement learning framework for optimizing cooling control of modern data centers to meet the challenge of large scale and safe AI deployment.

Abstract: Data centers power the internet, making digital communication and connection possible. In 2022, 1-1.3% of the total energy consumption in the world went to data centers. They also consume a lot of water in their cooling systems. Optimizing their energy and water usage is a priority in the industry. This paper presents one of our approaches to optimize water and energy in data center cooling by leveraging a Simulator Based Reinforcement Learning method. First we developed a physics-based simulation model that can predict the thermal behavior within 1°F of MAE (mean absolute error) for cold aisles. Then an RL model is trained offline resulting in a better policy for controlling the supply airflow setpoint. The production model at one of our data center regions has shown reduction of supply fan energy consumption by 20% and water usage by 4% on average across various weather conditions.

Submission Number: 7