Statistical analysis for predicting location-specific data center PUE and its improvement potential

This paper presents a statistical framework for predictive analysis of data center power usage effectiveness (PUE), with a focus on hyperscale data centers (HDCs). Thermodynamics-based PUE models considering representative economizer choices are proposed, taking climate variables and energy system parameters as inputs for robust PUE predictions. Sobol's method is used to assess total order sensitivity indices of key modeling parameters, suggesting that climate variables and uninterruptible power supply (UPS) efficiencies are the most important parameters. The PUE values of 17 HDCs operated by Google and Facebook were predicted, considering location-specific weather conditions, and uncertainties in energy system parameters and economizer choices. Results were verified using reported PUE values, indicating the model's effectiveness in capturing regional and seasonal PUE variations, and in generating point-estimations for macro-level data center (DC) energy models. Finally, achievable PUE values were computed through differential evolution, identifying minimum practical PUE values that could be obtained with state-of-the-art technologies. The framework can be applied in predictions of location-specific PUE values, PUE improvement analysis, and PUE target-setting by policy makers.