Physics-Guided Multi-Output Surrogate Modeling of a Centralized-UPS Data Center
Abstract
Data centers supplied through centralized uninterruptible power supplies (UPSs) are converter-dominated loads whose grid response reflects IT demand, power-conversion controls, energy storage, cooling, and site-support dynamics. This paper develops and electrically realizes a causal multi-output surrogate for a nominal 1-MW, 0.48-kV simulated data center. Four physics-informed terminal features-normalized IT demand, positive-sequence voltage deviation, voltage-unbalance ratio, and frequency deviation-are used to predict active and reactive power at both the point of common coupling (PCC) and UPS output. Six causal surrogate families are evaluated using separate training, validation, and held-out trajectories. A ridge-regularized low-order nonlinear autoregressive model with exogenous inputs (NARX) achieves the lowest validation and test errors, with mean four-output errors of 0.1959\% and 0.1441\% on a 1-MVA base, respectively. The selected surrogate is then coupled to a three-phase electrical interface in MATLAB/Simulink, where predicted PCC active power controls a variable resistive load and predicted reactive power is realized through controlled current sources synchronized to the PCC voltage. Across two 41-s AI workload profiles, including an additional pipeline parallelism trajectory, the electrical realization tracks the NARX commands with active- and reactive-power RMSEs of 0.604 kW and 0.353 kVAR for Evaluation Profile A and 1.539 kW and 0.442 kVAR for Evaluation Profile B, respectively. The results demonstrate a compact terminal surrogate that can reproduce both predicted power trajectories and their corresponding three-phase electrical behavior under measured PCC excitation.
- Comment
- Submitted to IEEE PES General Meeting 2027