Physics-Informed Predictive Causality in Data Center Cooling

Understanding causal structures in data center cooling systems is essential for energy optimization and fault detection. Conventional methods based on physical connectivity ensure interpretability but often miss latent interactions, while Granger causality captures predictive dependencies yet suffers from sensitivity to data quality and ambiguous directionality. To overcome these limitations, we propose a hybrid causal discovery framework that integrates physics-informed priors with Granger-inspired predictive modeling. A key innovation is the use of a relative increment formulation, which focuses on the proportional change in observables immediately after control actions. This design filters out long-term seasonal trends and emphasizes short-term, actionable effects. Applied to a large-scale dataset from a real data center, the framework successfully recovers known control–feedback links, identifies consistent control–temperature relationships, and reveals cross-unit influences overlooked by traditional approaches. By combining physical priors for directionality with predictive causality for flexibility, the method yields a causal network that is both interpretable and robust, offering a principled basis for decision-making in energy-critical infrastructures.