Dennis Hertz has been staring at load curves for eleven years, but he's never seen anything quite like the graph on his monitor from June 24, 2025. The line representing actual electricity demand shoots upward past his forecast like a middle finger to mathematical modeling. Five percent above prediction. In grid operations, that's not a rounding error. That's a near-miss with cascading blackouts across thirteen states.

We're meeting in a conference room at PJM Interconnection's control center in Valley Forge, Pennsylvania, three weeks after the heat wave that pushed the Mid-Atlantic and Midwest grid to 161 gigawatts, requiring nearly 4,000 megawatts of demand response to keep the lights on.¹ Hertz is a load forecaster, which means his job is predicting how much electricity people will use tomorrow, next week, next summer. It used to be about sophisticated pattern recognition. Now it's about watching those patterns dissolve in real-time.

He's 43, wearing a polo shirt with the PJM logo, and he has the particular exhaustion of someone who's been proven wrong in a very public, very consequential way. Not wrong because he's bad at his job. Wrong because the fundamental assumptions underlying his models are becoming obsolete faster than he can update them.

"I should clarify," he says before we even start, "that Dennis Hertz isn't my real name, and I'm a composite of several people who do this work. But everything I'm going to tell you is real. The data is real. The models breaking down are real. The existential dread is definitely real."

So let's start with June 24. You forecasted 154 gigawatts for the summer peak. You hit 161. What happened?

Dennis: What happened is that every single variable in my model decided to be wrong in the same direction at the same time. Variables are supposed to distribute randomly around your prediction. Some high, some low, it averages out. But when it's 102 degrees with 70% humidity in Philadelphia and 98 in Chicago and 105 in Columbus all at once, and everyone's AC is running at maximum and the solar panels are underperforming because they're too hot and three coal plants had to derate because their cooling water got too warm...

He stops, laughs. It's not a happy laugh.

Dennis: Sorry. What happened is climate change. That's the answer. But I can't put "climate change" into my forecasting model as a variable. I need specific, quantifiable inputs. And those inputs are changing faster than I can recalibrate.

Walk me through how you build a forecast.

Dennis: Okay. Traditionally, load forecasting is about three things: weather, economics, and historical patterns.

Weather is the biggest driver. Temperature, humidity, cloud cover. I use weather service forecasts and convert them to expected electricity demand based on historical relationships. Like, historically, when it hits 95 degrees in Baltimore, we see X amount of load. When it hits 100, we see Y amount.

Economics is industrial activity, employment, GDP growth. Are factories running? Are data centers expanding? That stuff.

And historical patterns are... this is where it gets interesting. Basically, people are creatures of habit. They use electricity in predictable patterns. Monday through Friday looks different from weekends. July looks different from October. You build models based on years of data showing these patterns.

And now?

Dennis: And now the historical relationships are degrading. That's the technical term. Very polite. What it means is: the past is becoming a worse and worse predictor of the future.

Take the temperature-load relationship. For decades, you could plot temperature against load and get a nice curve. Every degree above 85, load increases by roughly this much. Very reliable. But starting around 2022, 2023, that curve started shifting. Same temperature, higher load.

Why? Partly because people are running AC more. They're more aware of heat danger, they're more anxious, they're compensating. Partly because buildings that never had AC are installing it. Partly because the AC units themselves are working harder in higher humidity.

But here's the thing: those factors don't change linearly. They change in these weird, lumpy ways that are really hard to model. And they're changing while I'm trying to forecast.

You said your summer outlook forecasted 154 gigawatts. But you'd seen demand growing. Ten gigawatts between summer 2024 and 2025, more than double the previous year's growth.² Why didn't that show up in your forecast?

Dennis: Oh, it did. That's the really fun part.

Our forecast included aggressive growth assumptions. We knew demand was accelerating. We built that in. And we were still five percent low.

You want to know what keeps me up at night? It's not that I missed by five percent. It's that I don't know if next summer I should assume another 10 gigawatt increase, or 15, or 20. The trend line is bending upward, but I don't know the shape of the curve. I'm extrapolating from two data points and trying to predict 2026, 2027, 2030.

And meanwhile, we're retiring dispatchable generation. MISO alone lost 1,575 megawatts of generation that you can turn on when you need it.³ So I'm forecasting higher demand and lower reliable supply.

He makes a gesture that I interpret as "the exact opposite of what you want."

What does "dispatchable" mean in practice?

Dennis: It means I can call them up and say "we need more power right now" and they can deliver it. Coal plants, gas plants, some hydro. Nuclear runs all the time, you can't really dispatch it up and down quickly. Wind and solar are weather-dependent. You get what you get.

The grid is retiring a lot of coal plants because they're old and expensive and dirty. Which, like, I get it. Climate change is bad. But from a grid reliability perspective, I'm watching my margin for error shrink while my error rate goes up. Those two trends intersecting is not great.

How much margin do you have?

Dennis: Depends on the region. MISO is at "elevated risk" of supply shortfalls, which is one step below "high risk," which is one step below "we're gonna have rolling blackouts." PJM is better but not by much. We can provide 179 gigawatts today. Our forecast for 2035 is 210 gigawatts.⁴ That's a 31 gigawatt gap in ten years.

And that's assuming normal growth patterns. If we get a decade of accelerating heat waves like we got in 2024-2025, that 210 number is low.

You're describing a situation where you're always going to be wrong.

Dennis: I'm describing a situation where the fundamental premise of my job—that the future resembles the past in predictable ways—is becoming false. And I still have to do my job.

There's this thing that happens in forecasting called model drift. Your model slowly becomes less accurate over time as conditions change, so you periodically retrain it on new data. Normal model drift happens over years. I'm seeing it over months now. Sometimes weeks.

In June, we had a stretch where I was updating my next-day forecast three times because the weather kept getting worse and my load predictions kept being low. At a certain point, you're not forecasting anymore. You're just narrating what's happening with a 12-hour delay.

What does that feel like?

Dennis: Honestly? It feels like being a climate scientist. Which is not what I signed up for. I signed up to do electrical engineering and applied statistics. But now I'm in this weird position where I'm watching climate acceleration in real-time through electricity demand data, and I'm one of the first people to see it because load curves are incredibly sensitive to behavior change.

Like, I can tell you that people are running their AC more anxiously than they did five years ago. I can see it in the data. The evening peak is extending later. The overnight minimum load is higher. People aren't turning off their AC at night anymore. That's not a weather change. That's a psychological change in response to climate.

And I'm supposed to quantify that and build it into my models, but how do you model anxiety? How do you model the collective decision of millions of people to prioritize comfort over electricity bills because they're scared of heat?

Are you scared of heat?

Dennis: I installed a whole-home generator last year. Make of that what you will.

PJM activated 4,000 megawatts of demand response during the June heat wave. What does that actually mean?

Dennis: It means we paid industrial customers and some large commercial buildings to reduce their electricity use during peak hours. Factories slow down production lines. Data centers shift computing loads. Big office buildings raise their thermostat setpoints a few degrees.

It works, mostly. But it's expensive, and it's not unlimited. We have about 8 gigawatts of demand response capacity total.⁵ We used half of it in one afternoon. And that was with all our generation running at maximum, including plants we'd recalled from maintenance.

The grid operator, the people in the control room making real-time decisions, they're basically playing Tetris with gigawatts. They're trying to fit available supply to actual demand, and they have very few pieces to work with, and the pieces keep changing shape, and if they fail, people die. Hospitals lose power. Dialysis centers go dark. Nursing homes can't run AC.

How do you live with that?

Dennis: I update my models. I write reports explaining what went wrong. I build in more conservative assumptions. And I watch the gap between my forecasts and reality keep growing anyway.

There's a line I keep thinking about from this energy consultant, Alison Silverstein. She said we're asking the grid to do "heavy walking, if not flat-out sprinting, for more and more hours and days of the year."⁶ That's exactly right. We built this infrastructure for a world where peak demand happened maybe 100 hours a year. Now it's 200 hours. Soon it'll be 300.

And I'm the person who's supposed to predict which 300 hours. With models built on data from when it was 100 hours.

What happens in five years?

Dennis: In five years? Honestly, I think my job looks really different.

Either we have much better real-time forecasting, like AI models that can adapt faster than I can, or we have much more distributed generation and storage so that load forecasting matters less because the grid is more resilient to misses.

Or we have rolling blackouts during heat waves and I'm the guy who gets blamed for not predicting it, even though everyone in this building knows the infrastructure can't keep up with demand growth.

The thing nobody wants to say out loud is: we might need to have a conversation about demand management that's not voluntary. About saying "you can't run your AC below 78 degrees during peak hours" or "data centers need to be in places with reliable cooling, not just cheap land." But that's politically impossible, so instead we just keep forecasting and hoping we build enough generation in time.

Do you think we will?

Dennis: Build enough generation? I think we'll build some. I don't think it'll be enough. The economics are complicated, the permitting is slow, and the demand growth is faster than anyone expected.

What I think actually happens is: we have a major grid failure in the next three to five years. Something bigger than the Texas freeze, bigger than the June heat wave. Maybe PJM, maybe ERCOT, maybe MISO. And then we have the political will to make big infrastructure investments, and we do a bunch of things we should have done ten years ago.

And my forecasts will still be wrong, because the climate is changing faster than we can adapt. But maybe the grid will be resilient enough that being wrong doesn't kill people.

He closes his laptop. On the screen, I catch a glimpse of a graph before it goes dark. A jagged red line climbing upward, chasing a blue line that's always just out of reach.

Footnotes

1. https://power.mhi.com/regions/amer/insights/summer-grid-reliability-2025 ↩

2. https://epsa.org/summer-2025-outlook-rising-temperatures-tightening-power-supply/ ↩

3. https://epsa.org/summer-2025-outlook-rising-temperatures-tightening-power-supply/ ↩

4. https://www.technologyreview.com/2025/06/26/1119358/summer-grid-ai-air-conditioning/ ↩

5. https://www.technologyreview.com/2025/06/26/1119358/summer-grid-ai-air-conditioning/ ↩

6. https://www.npr.org/2025/06/26/nx-s1-5443660/amid-extreme-heat-some-power-grids-may-struggle-to-keep-up-with-rising-energy-demand ↩