One DMD Mode Explains 93.6% of AI Data Center Load Correlation

Hyperscale AI data centers produce load fluctuations so tightly coupled that a single Dynamic Mode Decomposition eigenvalue at 0.005 Hz captures 93.6% of the correlation energy across a 39-bus power system.

Why Classical Independence Assumptions Fail for AI Data Centers

Standard power system analysis assumes loads fluctuate independently or with simple time-averaged spectra. Hyperscale AI data centers shatter that assumption. Their workload-driven power draws create spatially and temporally correlated fluctuations that are episodic and non-stationary. A new modal analysis applies Dynamic Mode Decomposition (DMD) to the temporal evolution of pairwise inter-bus correlation coefficients on an IEEE 39-bus Real-Time Digital Simulator (RTDS) testbed with three converter-interfaced AI data center loads driven by synthetic workload profiles.

DMD eigenvalues encode the correlation regime directly: sustained coherence sits near the unit circle, decaying transients fall inside, and intensifying events push eigenvalues outside. A global, time-averaged DMD identifies a slow thermal band at $f \approx 0.005$ Hz with eigenvalue magnitude $|\mu| = 0.91$ that alone captures 93.6% of the total correlation energy. That single mode dominates the system's load correlation structure.

Modal Analysis Reveals Transient Intensification Windows

Global averaging hides transient events. A sliding-window DMD formulation catches them. Out of 775 windows analyzed, 51 (6.6%) satisfy the $|\mu_k^{(n)}| > 1$ criterion, indicating eigenvalues outside the unit circle and correlation intensification. These windows align with stochastic workload coincidences among the three data center loads. Cross-validation with RTDS voltage coherence measurements confirms that coupling is significantly elevated during those intervals.

The paper's results show that the intensification events are not random noise. They match physical coupling mechanisms revealed by the oscillation frequencies. This means the correlation bursts are real, predictable in aggregate, and tied directly to how AI training and inference jobs overlap in time across multiple facilities.

An Early-Warning Signal for Grid Operators

The proposed modal growth indicator tracks the shift in eigenvalue magnitude across windows. It provides a signal that correlation is strengthening before the pairwise coherence peaks. For a power system that increasingly depends on a handful of hyperscale loads, that lead time matters.

This approach turns DMD from a post-mortem analysis tool into a preventive control signal. Operators could use it to trigger mitigation actions such as fast-ramping reserves or demand response before the correlation spike stresses voltage stability. The next step is clear: validate this modal growth indicator on real utility-scale data center load traces, then integrate it into real-time stability assessment platforms.

Source: Modal Analysis of Spatial Load Correlation in AI Data Center-Dominated Power Systems
Domain: arxiv.org