Transcript

Introduction

Good afternoon, everyone. Welcome to today’s white paper webinar. Today, I’m here with David Horning, and we’re gonna talk about data centers and grid readiness. Very important topic for utilities.

This paper is a summary of the grid readiness paper from the Industry Connections Report. And most of you have probably taken a call about data center interconnections and the problems with that in the last 12 months. The paper’s really about what your grid code is going to look like in the next 24 months, what you’ll expect to measure.

The 3 trillion global data center investment between 2026 and 2030, 3% of the global electricity today and up to 5% to 12% in three to five years is gonna be used for data centers. There’s a 250 terawatt per year forecast, which gives uncertainty of what to do, what to plan, what you will need.

And this is a much faster rate of connection than the utilities have ever had before, so there’s enormous financial pressure to get these data centers connected as quickly as possible regardless of the engineering behind that. It’s a challenge.

Why These Loads Are Different

AI Pulse Oscillations

You get your AI pulse oscillations, which is something totally new. You’ve got thousands of GPUs ramping up in lockstep on training steps and creates a periodic load oscillation. If the pulse frequency lines up with the transmission resonance or capacitor bank, you’ll get amplified voltage instability.

This is not a classical harmonic distortion. It’s a load-driven, and the frequency depends on the model being trained. As the model changes, so does the load. You’d get hundreds of milliwatts changes in seconds, and you get power swings of hundreds of milliwatts, and there was a 1500 megawatt voltage-sensitive load reduction in the Texas.

As the UPSs and the power supply rectifiers, if you have a low sort-circuit ratio, you can get the oscillations from the load. The harmonic resonance is amplified by the dataset or via capacitor banks.

Large Load Swings

This is a real problem in a couple different ways. First, large load swings where you might have 10, 20, 100 megawatt load steps as the entire data center or bank of GPUs in the data center finishes an AI training task or inference task. And so you have these load steps that are large and discreet on a second basis or minute spaces that traditional loads don’t have.

You don’t have the load diversity inside data centers you would have in traditional giant loads, because all those GPUs are tied together doing the same process, which can stop or start on a millisecond level timeframe.

UPS Interactions and Sub-Synchronous Oscillations

Another challenge are the UPSs that control these. These data centers generally are run powered by UPSs. The utility load is an AC to DC converter of the EPS, and the UPSs, which are huge UPSs, are powering the data center servers themselves.

Those UPSs can interact with each other. The GPSs have internal control loops to keep their output voltage regulated as they themselves see these giant load swings. And those loop frequencies tend to be in the tens of hertz range. And so you’ll get these low frequency oscillations.

Those SSO, the sub-synchronous oscillations, not necessarily from the load, the GPUs themselves, but from the UPSs that are powering those GPUs. Very similar to solar inverters, where firmware in the inverter or control loop kind of dictate how well-behaved they are. Same thing for the GPSs. They can be parallel, they can be segregated, they can interact with each other.

These cause all sorts of variations on the current that they draw from the utility, and this shows up as these sub-sequence oscillations, and these can be amplified by slow frequency system resonances, and these can cause trouble in a couple different ways.

Generator Mechanical Resonance and Flicker

One of the most concerning for transmission-connected data centers that are 100 megawatt or even gigawatt ranges, hyperscaler systems, are the fact that those 10 to 20 hertz oscillations can be right in the sweet spot of mechanical resonance on generators. There’s a lot of concern about generator shafts vibrating or mechanical instabilities from generators when the loads are pulsing at those resonant frequencies. There’s a lot of concern about how to monitor that, mechanical vibrations in the chat.

So that’s one area where these sub-sequence oscillations cause trouble on the transmission side, and they cause trouble in distribution by manifesting itself as flicker. So if a data center is in a distribution network, it’d be 10 or 20, 30 megawatts and still cause lots of flicker for many customers.

This can cause GFCI breakers to trip. The sub-sequence oscillations in conjunction with six Europe’s harmonics create sidebands around those frequencies. So, for example, the fifth harmonic at 300 hertz, you can generate sidebands around the fifth harmonic, inner harmonic frequencies that manifest themselves as flicker or sensitive points to breakers.

It’s a whole new category of load behavior that causes voltage distortion in ways that the standards aren’t designed to handle because you didn’t really have loads doing this in the past.

Example: Monitoring a Data Center with a Seeker

An example that David talks about here in the paper: here is an example where we have a seeker monitoring a data center. On the left side is a waveform capture. This is a five-minute waveform capture, and this seeker is set up to look at a waveform every five minutes, and this is a five-second long waveform capture.

If you just look at the sine waves on a five-second basis, nothing looks unusual. You zoom in, voltage is in red, current is in blue. Nothing really looks that bad. You see some current distortion but that’s not terrible. On a first glance, looking at the waveforms, you don’t see anything wrong.

If you look at the RMS version of this though, if you click on RMS, here the software is creating a continuous RMS graph. This is a one cycle wide, long window that moves one sample at a time showing the RMS value, and you can already see in blue these periods where over five seconds the current rises and falls on a periodic basis.

But if you zoom in, you can see a faster period. Now you see the current going up and down, up and down, up and down, and this is about 5 amps of oscillation on the primary side, on a 34 KV primary. And then if you zoom in again, you can see yet a third oscillation. This is higher frequency oscillation on top of that slower one.

So we really have three different frequencies going on here. We have a few hertz, and we have about 20 hertz, and we have one even faster, all in this one five-second capture, and this changes over time.

On the graph on the right, we have a single-cycle RF chart, and you can see it going bursty where it goes in the bursts of going up and down and then it’s steady again. And you can even look at this on a spectrogram basis where in this plot, we’re looking at a spectrogram over time, over several days of a data center. On the left where it’s kind of the bright orange-yellow, that’s a 20 hertz to 30 hertz oscillation, and then the data center switched to a different mode with the GPSs and that went away for several days.

So a lot of this is within the control of those GPSs, depends on how they operate and what the data center is doing. These are some of the things you’ll see in data centers that can really cause trouble in distribution and in transmission.

Regulatory Response

Another problem is that data centers can be built in months while the generation and transmission upgrades take years. By the time the study is done, you’re already behind on the next one or two data centers coming. That’s why the regulators are moving fast.

There’s been announced regulations, or people coming up with regulations. We have the FERC coming up with the 20 megawatt large loads. Your FERC is coming up with the generation style interconnections 60-day curtailable timelines. This is just for generator style interconnection procedures for loads including expected 60-day timelines for curtailable loads.

The NRC PFI reliance guidelines identified seven areas of risk including PQ and cybersecurity, and for Q2Q26, we’ll have standardized how data centers get modeled for power systems study that affect your interconnection studies.

Common Requirements Across Grid Operators

They’ve done studies for over three continents and utilities who’ve come up with their own standards and they’ve come up with a set of common requirements including:

• Voltage frequency ride through
• Ramp rate limits
• Oscillation control
• Power factor and harmonics
• Post-fault active power recovery
• Validated dynamic load models

This was a survey of 12 grid operators across North America, Europe, and Australia, and they’re finding the same areas are needed in the short list of the areas of concern.

Computing hardware is designed to a single global standard, yet must comply with dozens of divergent grid codes. Fragmentation lands on the utility, every interconnect becomes a negotiation. Until there’s standards, you have to negotiate each connection, which is why the standards are needed to pull all this together into a set of standards so we don’t have to negotiate and spend time figuring out each one independently.

Upcoming Standards

They’re gonna have standards for:

• Interconnection processes and modeling data
• Performance interoperability
• Ride through, ramp, oscillation, and post-fault recovery
• Co-located generation coordination
• Certification and testing
• Chips to grid design philosophy

Some of it will fall on the semiconductor industry to meet certain standards to be in data centers.

Measurement Requirements and PMI Solutions

When you’re going to monitor these for power quality things, you need to capture:

• Subcycle voltage transients
• D during the AI training cycles for your ramp limits
• Low frequency oscillation content
• Harmonic spectrum under bursty loads
• Post-fault active power recovery characterization

These are measurement problems, and since they’re new, have to find new ways of capturing new things to look for. The PMI Recorders can capture this and alert on problems, and our MerlinAI can help focus on the problems.

We are developing Merlin AI changes for data centers so we can provide better information for data centers. As we capture the data, we can come up with better models to highlight this.

The standards aren’t harmonized yet, but the measurement requirements are landing on your desk now. And again, this is a summary of the IEEE IC25004, and there’s a link to the document that this white paper summarized. Thank you for attending. If you have any questions, feel free to call.