Data-Driven Cooling Energy Prediction Across AI Workload Intensities: Statistical and Regression-Based Comparison of AHU and Chiller Cooling Systems

As data centres become the backbone of global digital infrastructure, understanding the energy dynamics of their cooling systems is critical to achieving sustainable computing at scale. The rapid expansion of artificial intelligence (AI) workloads has substantially increased the energy and thermal management demands of modern data centres. This study presents a data-driven statistical analysis and predictive modelling comparison of two prevalent cooling system architectures — air handling unit (AHU)-dominant and chiller-dominant systems — using a merged operational dataset of over 2,500 records drawn from publicly available Kaggle repositories and the Frontier high-performance computing (HPC) system. All energy and power values in the analysis are expressed in kilowatts (kW) to ensure consistency across merged sources.

Pearson correlation analysis, Welch’s two-sample t-tests, Cohen’s d effect size estimation, a Kolmogorov–Smirnov (KS) distributional check, and multivariate ordinary least-squares (OLS) linear regression were applied to examine associations among server workload intensity, cooling system classification, ambient outdoor temperature, and cooling unit power consumption. Results indicate that server workload percentage is the variable most strongly associated with cooling energy demand (bivariate R² = 0.95), while ambient outdoor temperature shows negligible linear correlation with indoor cooling requirements (r ≈ 0). Chiller-classified records were associated with mean cooling energy approximately 20.9% higher than AHU-classified records (419.8 kW vs. 347.8 kW; p < 0.001; Cohen’s d = −2.534, large effect). A three-feature multivariate regression model achieved R² = 0.954 and MAE = 7.87 kW on a held-out test set, matching a workload-only baseline, which reinforces that server workload is the dominant predictive signal. All findings are associative and observational in nature; no causal claims are made.