Pacific Northwest National Laboratory Says Phasor Measurements Can Understate Higher-Frequency Oscillations from AI Training Data Centers

Pacific Northwest National Laboratory issued its report PNNL-39191 in April 2026. It was prepared for the Energy Department. The report evaluates whether existing measurement systems are adequate for monitoring the oscillations that large data center loads can put on the grid. Its short answer is that phasor measurement units, which form "the backbone of existing wide-area monitoring systems," track the slow swings well and can misread the faster ones. It then ties that to limits that utilities and system operators have proposed on how much a data center's load may swing. Parts of the report follow, not the whole.

A Load With a Rhythm

On the computing side, the report says AI training facilities rely on large numbers of graphics processors "operating in parallel during tightly synchronized distributed training cycles." Each training iteration alternates between a high-power phase, when the processors compute, and a lower-power phase dominated by memory transfers, communication and synchronization. Repeated, that alternation "produces a sustained oscillatory load profile."

Conventional data centers, it says, "generally exhibit relatively smooth and diversified demand profiles." Their workload consists of many independent tasks whose fluctuations "tend to average out over time." In a training cluster, coordinated work "can cause power demand to rise and fall simultaneously across many servers."

The swing has more than one speed. The iteration cycle gives a slower component, and variation inside the computing phases gives faster ones that may reach the sub-synchronous range, which the report places, for a 60 Hz system, generally between about 5 Hz and 59 Hz. Oscillations "may also emerge from non-AI conventional data centers," the report adds, typically from their power electronics and "control interactions rather than cyclical computational demand."

What the Swing Can Reach

What the swing can do to the grid depends on its frequency, the report says. At low frequencies, periodic load variations "may interact with poorly damped inter-area and regional oscillation modes." If a load's forcing frequency is close to a system mode, the injected power "can amplify the existing oscillation through resonance, potentially increasing power swings across major transmission corridors and tie-lines." At higher frequencies, the load can interact with the torsional modes of nearby synchronous generators, which the report ties to the mechanical shaft dynamics of turbines and generators. If those modes are excited repeatedly, "cyclic torque can accumulate in generator shafts, potentially leading to increased mechanical stress and long-term fatigue."

Historically, the report says, such torsional interactions "have been short-lived events that subsided once mitigation actions were implemented." Oscillations from tightly synchronized computing in large AI data center clusters "may persist continuously during normal operation of the facility."

What the Meters Underestimate

Phasor measurement units were developed to track slow electromechanical swings, and the report says two decades of field use have made them "a trusted foundation for wide-area oscillation monitoring." At higher frequencies, the report names two main factors that limit them. Their filters are "intentionally designed to be selective around the nominal fundamental frequency," which leads to "systematic underestimation of higher-frequency oscillation magnitudes." And oscillations above half the typical reporting rate "cannot be faithfully represented and may appear at incorrect frequencies due to aliasing."

A recent field event, the report says, "illustrates these effects." In October 2024, ERCOT observed power oscillations tied to ramping at "a large crypto-mining data center load," and the report's figures of the event are credited to ERCOT. Phasor data reported 30 times a second from the point of interconnection showed an oscillation of about 7.5 Hz, with a peak-to-peak magnitude of 25 MW. Waveform data from a digital fault recorder at the same location showed that "the actual oscillation" was about 23 Hz and 50 MW peak to peak. The report puts the difference down to the phasor unit's limited reporting rate, which caused "aliasing of the oscillation frequency and significant attenuation of the measured magnitude." Same location, same event. Two pictures.

Under the IEC/IEEE standard the report applies, the dynamic performance of these units is validated primarily for swings up to about 2 Hz in P-class devices and about 5 Hz in M-class devices. As a result, the report concludes, they "are generally not well suited for monitoring oscillations significantly higher than approximately 5 Hz."

Proposed Limits

The report summarizes representative examples of the "preliminary compliance criteria" that several utilities and system operators in North America have proposed. In its summary of ERCOT's, load power "must not repetitively exceed a 10 MW change within a 5-second sliding time window." The criterion it attributes to LIPA, on Long Island, is defined on power derived from waveform measurements. In its high band, "the power summed over any two adjacent frequency bins, computed using a 1-second FFT window and averaged over a 10-second rolling window, must not exceed 3.5 MW."

To illustrate the effect, the report builds an example on that criterion. It synthesizes a 20 Hz oscillation with a true amplitude of 4.8 MW, above the threshold. Measured from the waveform, the oscillation is detected. Run through phasor estimation at 120 reports a second, the estimated magnitude "converges to a value that remains below the detection threshold." The report states the limits of its own example: the source of LIPA's criterion "does not allow for the use of PMU data," the phasor version "is for illustration only," and the results "are not intended to evaluate" LIPA's method. At 1 Hz, the two measurements give essentially the same result.

When compliance assessments rely solely on phasor measurements, the conclusions say, "a critical oscillation could be incorrectly assessed as benign, potentially delaying mitigation actions or obscuring the true dynamic impact of the load." Correcting the readings is "challenging" in practice, the report says, because in most commercially deployed units "the phasor estimation algorithms and associated filter designs are proprietary." A unit's attenuation may be measured in a laboratory instead, it suggests, giving approximate correction factors.

A site that looks within its limit on phasor data could be over it on the waveform.

Waveforms, and Their Practical Challenges

The report's answer is to add high-resolution waveform measurement, which "can capture oscillations across a broader frequency range." Continuous waveform monitoring, it adds, "introduces practical challenges related to large data volumes, communication bandwidth, and real-time data processing." Phasor units stay for the slow swings in its account, with waveform measurement beside them, and it calls hybrid approaches "a promising pathway." Waveform data, it notes, "are not universally available across transmission networks."