Temporal causal feature selection for robust machine learning modelling of data center operations

Modern data centers are complex, energy-intensive systems whose efficient operation is critical to our global digital infrastructure. This paper introduces the TempCaFe (Temporal Causal Features) framework that applies causal feature selection to enhance predictive modelling and intervention analysis in data center environments. By integrating causal discovery with domain knowledge for time-series data, the approach not only improves the predictive performance of machine learning models but also their robustness and interpretability. We evaluate the utility of the proposed framework by training 3,660 models for 31 prediction tasks with data from a data center simulation, a physical testbed, and two operational data centers with distinct cooling architectures (air and water). The results indicate that causal feature selection enhances prediction accuracy and leads to significantly better results after system interventions, outperforming traditional feature selection methods and using all available features. Furthermore, models trained with causally selected features exhibit greater robustness to data scarcity (number of samples) and require less than one-third of the original features. These findings underscore the potential of causal machine learning to enable robust modelling of data center operations and serve as a foundation for decision-making, such as what-if analysis and design of control systems.