DMD method catches AI data center load fluctuations with 93.6% accuracy

Hyperscale AI data centers—with their massive, bursty workloads—induce spatially and temporally correlated load fluctuations that violate classical power system independence assumptions. Traditional time-averaged spectral methods fail to capture these non-stationary, episodic correlations. This paper from Chaudhary et al. applies Dynamic Mode Decomposition (DMD) to the temporal evolution of pairwise inter-bus correlation coefficients, forming a low-dimensional state representation without requiring stationarity. Using an IEEE 39-bus Real-Time Digital Simulator (RTDS) testbed with three converter-interfaced AI data center loads driven by synthetic workload profiles, the researchers demonstrate that global DMD provides a baseline modal analysis: a slow thermal band at ~0.005 Hz with eigenvalue magnitude 0.91 captures 93.6% of total correlation energy.

A sliding-window DMD formulation then identifies transient intensification events: 51 of 775 windows (6.6%) satisfy the |μ| > 1 instability criterion, aligning with stochastic workload coincidences. Cross-validation with RTDS voltage coherence confirms elevated coupling during these intervals. The proposed modal growth indicator serves as an early-warning signal before peak pairwise coherence, enabling grid operators to anticipate power quality issues. This work directly addresses a growing challenge—how to maintain stability as AI computing loads become a dominant share of energy consumption.