Picture a data center in the MISO footprint drawing 300 megawatts (MW). An on-site gas generator is covering 100 MW of that, so from the grid’s point of view, the facility is only pulling 200.
Then the generator trips.
In an instant, MISO is serving all 300 MW, with nothing in its forecast to warn it. I walked through this scenario in PCI’s Oct. 15 webinar, “168 Hours to 9 Minutes: MISO’s New Large Load Forecast Requirements,” because it gets at why MISO is changing who forecasts large loads in the first place.
Following FERC’s June 2026 show cause order, MISO filed new rules on Aug. 28 that will require transmission customers serving a large load to submit their own forecasts for it, hourly and every five minutes, starting Sept. 1, 2027. In that filing, MISO acknowledges that these loads can’t be accurately forecast with the historical, weather, and calendar data its own short-term system relies on. For the utilities serving data centers, a forecasting job MISO has always handled is about to land on their desks.
Why one data center is harder to forecast than a whole city
Utility load forecasting works because it forecasts crowds. A city’s demand is the sum of thousands of homes and businesses, and their small individual errors tend to cancel each other out. What’s left is a smooth curve that follows temperature, time of day, day of the week, and holidays closely enough to model well.
A single data center breaks those assumptions. Its demand follows the work happening inside it: when an AI training job starts, when new servers come online, when equipment goes down for maintenance. A large training job can run near full power for days, then drop sharply the moment it finishes. Instead of smooth curves, you get steps and plateaus, and there’s no crowd for the errors to hide in.
Not every large load is that unpredictable. Cloud storage tends to run fairly flat, since it’s mostly reading and writing data. But AI and high-performance computing loads are dynamic and driven by workload, which makes them the hardest to forecast of any large load type MISO describes.
You can’t forecast a training run from weather. The best input for a data center forecast is the operator’s own schedule, and without it, a forecaster is working blind.
What MISO now requires, and from whom
Under MISO’s filing, the obligation falls on the transmission customer, meaning the company that buys transmission service to serve the large load. That’s usually a utility, not the data center itself, which means the company accountable to MISO for forecast accuracy often isn’t the one whose operations drive the load.
For each large load it serves, that transmission customer will owe MISO two forecasts:
- An hourly forecast covering the next 168 hours, refreshed at least every hour
- A five-minute forecast covering the next six hours (72 intervals), submitted nine minutes ahead of each interval
Each large load also gets its own Commercial Pricing Node (CPNode), separate from the load zone a utility already bids into, and its forecasts are submitted at that node.
The requirement that tends to surprise people is that MISO wants gross forecasts. In the opening example, that means submitting the full 300 MW the data center consumes, not the 200 MW it draws from the grid. If the generator trips, the grid sees the whole load at once, and a netted forecast would have hidden that risk. My practical advice is to forecast the facility’s consumption as a single number rather than modeling the load and the generator separately, since two forecasts means two sets of potential errors.
Dwarak Varadharajan, PCI’s Senior Product Manager of Optimization & AI, on why MISO requires gross forecasts for large loads with on-site generation, from PCI’s Oct. 15 webinar, which is available to view upon request.
The forecast that looks fine but isn’t
I like to compare one of the biggest risks to a GPS that loses its signal. The screen still shows your position, so everything looks normal, but it’s stuck on the last place it knew you were.
Forecast submissions can fail the same way. If MISO rejects a submission, it falls back on the last forecast it accepted, so nothing looks broken while MISO quietly works from stale numbers. If nothing is submitted at all, MISO falls back on the last measured load, which knows nothing about when the next training job starts.
Either way, forecast error has a cost. MISO compares submitted forecasts against actual telemetry, and as my colleague Khai Le, PCI’s Senior Vice President, pointed out in the webinar, deviation charges and real-time price volatility already turn load forecast error into real money in RTO markets today. MISO is also exploring a new deviation charge specifically for large loads, though that framework is still being designed.
The volume alone rules out doing this by hand. A five-minute forecast for one large load means 288 submissions a day, and a utility serving several large loads multiplies that. Automation here is a necessity.
Forecasting a load that runs on a schedule
If the operator’s schedule is the best input, the first job is getting it. A reliable large load forecast starts with a way for the load customer to share its plans: when training runs will start, when new servers come online, what cooling they’ll need, and when maintenance or agreed curtailments are coming. Those updates have to arrive quickly, since an out-of-date schedule is useless for a six-hour or five-minute horizon. For developers, I’d recommend building that data sharing into service agreements from the start.
The modeling itself takes more than one approach. When an attendee asked whether statistical models are enough, my answer was that we use both. At PCI, we pair physics-based models, which calculate server power requirements, with machine learning, which picks up cooling demand that rises and falls with ambient temperature.
Learn more about how PCI Forecaster helps utilities automate forecasting for large loads.
Dwarak Varadharajan explains why PCI combines physics-based and machine learning models to forecast data center load, from PCI’s Oct. 15 webinar, which is available to view upon request.
Even when a third party supplies the forecast, the utility still needs a system to validate it, submit it to MISO, and shadow the settlement charges that follow. MISO holds the transmission customer accountable, not the data center, so that system has to run reliably around the clock.
PCI webinar: ‘MISO’s New Large Load Forecast Requirements’
The full recording goes well beyond what’s covered here. In it, you’ll find:
- A live demo of PCI Forecaster, including the data freshness indicators that flag a late or stale forecast and side-by-side benchmarking of multiple forecast sources
- MISO’s four tests for what counts as a large load, and how an expansion can turn an existing site into one
- The checkpoints where MISO scores your forecast, and the accuracy metrics it uses
- Khai Le’s walkthrough of the large load bid-to-bill workflow, from day-ahead demand bids to the monthly invoice
- The complete Q&A
If your team serves, or expects to serve, a large load in MISO, the recording is a practical place to start preparing for September 2027.
