Causal Machine Learning Approaches for Modelling Data Center Heat Recovery: A Physical Testbed Study

Data centers (DCs) form the backbone of our growing digital economy, but their rising energy demands pose challenges to our environment. At the same time, reusing waste heat from DCs also represents an opportunity, for example, for more sustainable heating of residential buildings. Modeling and optimizing these coupled and dynamic systems of heat generation and reuse is complex. On the one hand, physical simulations can be used to model these systems, but they are time-consuming to develop and run. Machine learning (ML), on the other hand, allows efficient data-driven modeling, but conventional correlation-based approaches struggle with the prediction of interventions and out-of-distribution generalization. Recent advances in causal ML, which combine principles from causal inference with flexible ML methods, are a promising approach for more robust predictions. Due to their focus on modeling interventions and cause-and-effect relationships, it is difficult to evaluate causal ML approaches rigorously. To address this challenge, we built a testbed of a miniature DC with an integrated waste heat network, equipped with sensors and actuators. This testbed allows conducting controlled experiments and automatic collection of realistic data, which can then be used to benchmark conventional and causal ML methods. Our experimental results highlight the strengths and weaknesses of each modeling approach, providing valuable insights on how to appropriately apply different types of machine learning to optimize data center operations and enhance their sustainability.