Data center cooling represents one of the fastest-growing and most variable electricity loads on modern grids, yet most power system models rely on simplified, static efficiency assumptions that fail to capture real thermal dynamics. This research addresses a critical gap by presenting a configurable simulation framework that generates realistic cooling electricity profiles for both air-cooled and liquid-cooled systems—the two dominant cooling architectures in large AI facilities.
The key innovation lies in modeling cooling load as a dynamic system rather than applying fixed coefficient-of-performance calculations. Traditional approaches assume constant efficiency ratios, but in reality, cooling demand fluctuates with data center workload, ambient temperature, humidity, and equipment operating states. These transients matter significantly for power system studies, which require accurate time-series load profiles spanning days or weeks.
Validation against operational data from the Marconi100 supercomputer demonstrates substantial improvements: mean absolute error dropped from 95.80 kW to 20.88 kW, while root-mean-square error fell from 109.79 to 27.27 kW. Across 520 daily profiles, the model faithfully reproduces both peak demand events and intraday variability patterns—essential for grid stability assessment.
For power system planners, this model bridges a knowledge gap at a critical moment. As artificial intelligence workloads proliferate, data center electricity consumption is projected to become a dominant grid driver in many regions. Accurate cooling load forecasting enables utilities to plan transmission capacity, schedule demand response programs, and assess reserve margins more reliably. The model's computational tractability makes it suitable for large-scale integration into power flow studies and operational planning tools.
The work underscores the importance of physics-based, configurable models over one-size-fits-all assumptions. As data centers increasingly employ hybrid and innovative cooling strategies, flexible simulation frameworks that adapt to facility-specific architectures will remain essential for grid reliability.
