Load forecasting: why DER signals break utility models

September 7, 2026 · 5 min · Free, with full transcript

Utility teams still treat load forecasting as a regression against temperature and industrial schedules. This episode argues that world is gone. Rooftop solar can flip the sign of the weather relationship, a passing cloud can return 40 megawatts in 90 seconds, and a midnight EV rate discount can synchronize 10,000 chargers into a spike that never existed before. Bolting DER signals onto the old model produces a more complicated wrong answer. You come away knowing where the money leaks, through NERC reliability reports and rate cases where a 2% error becomes a nine-figure conversation, why endogeneity breaks price-sensitive models, what probabilistic forecasting with confidence bands gives operators, and why dbt Labs style version control keeps models auditable without telling you they are wrong.

Listen to the podcast

5 min

Chapters

Key takeaways

  • List every rate change and demand response program launched in the last 18 months, then check whether your model has a variable that reacts to them.
  • Stop treating weather as a one-directional driver, since rooftop solar can cut afternoon demand instead of raising it.
  • Model EV charging as behavior shaped by your own price signals, not as a fixed 6 p.m. plug-in assumption.
  • Publish a range with confidence bands rather than a single megawatt number, so operators know when to keep a reserve plant warm.
  • Use pipeline tooling like dbt for auditability and speed of fixes, but cross-check vendor claims and treat model correctness as a human judgment.
Read the full transcript

Welcome to the Leaders Insights podcast.

Today's episode, load forecasting for utility operators integrating weather, behavioral

and DOR signals into demand prediction pipelines.

Everyone in your field keeps saying load forecasting is a solved problem.

You told me before we started, that's nonsense.

Go.

It was solved past tense.

For 40 years the utility playbook was, temperature goes up, air conditioners turn on, demand

climbs.

You regress against weather and industrial schedules, and you're within 2%.

That world is dead and a lot of forecasting teams haven't noticed the funeral.

Dead how?

Weather still drives demand.

It does.

But now half your grid fights back.

Rooftop solar means a hot sunny afternoon can cut demand instead of raising it.

The panels shave the peak.

Then a cloud rolls over a neighborhood and 40 megawatts reappears in 90 seconds.

Your old model sees weather and says, high demand.

Reality says, actually no.

The sign of the relationship flipped and nobody updated the equation.

So the fix is just adding solar data to the pipeline.

Problem solved again.

That's exactly the mistake everyone makes.

They bolt DR signals, distributed energy resources, meaning the solar panels, home

batteries, and EV chargers sitting behind the meter, onto the same regression and call

it modern.

What they've built is a more complicated wrong answer.

The behavior is the hard part, not the hardware.

Explain the behavior part.

Take EV charging.

Electric vehicle charging.

A utility in California assumes everyone plugs in at 6 p.m. when they get home.

Fine.

Then the utility offers a midnight rate discount.

And suddenly 10,000 cars all start charging at 12 hours, zero minutes, and one second because

they're on timers.

You've engineered a synchronized spike that never existed in nature.

Your forecast, trained on last year's behavior, has no idea the rate change even happened.

That sounds like the utility shooting itself in the foot.

Constantly.

The price signal changes the thing you're trying to predict.

Economists call that endogeneity.

The variable you're measuring reacts to your own decisions.

Most load models pretend the world holds still.

It doesn't.

Where does this actually cost money?

Give me the pain.

Two places.

NERC reliability reports.

That's the North American body that finds you when the lights flicker.

And rate cases where regulators decide what you're allowed to charge.

Under forecast the peak and you buy emergency power at 10 times the price.

Or worse, you shed load and end up on the news.

Over forecast and you've built a gas plant nobody needed.

And the regulator makes shareholders eat it.

Either direction, a 2% error, is now a nine-figure conversation.

The vendors say their tooling fixes this.

You buy that?

Partly.

DBB Labs, DWE, a data tooling vendor, publishes numbers on how much faster teams ship pipeline

changes with proper version control.

And it's genuinely useful for keeping these models auditable.

But they sell the software, so cross-check it.

The tooling helps you fix a model quickly.

It does not tell you the model is wrong in the first place.

That's a human problem wearing a software costume.

So what does the independent research say people should do differently?

MIT Technology Review had good reporting last year on probabilistic forecasting.

Instead of predicting one demand number, you predict a range with confidence bands.

As the honest answer on a partly cloudy day with high EV adoption isn't 4,200 megawatts.

It's somewhere between 3,800 and 4,600.

And I'm nervous.

Operators hate that.

They want a single number to plan against.

But a range is harder to act on.

Isn't the single number more useful?

A confident wrong number is the most expensive thing in the building.

The range tells you when to keep a reserve plant warm and when to relax.

That's the whole game.

Knowing your own uncertainty.

The dry truth in one line?

We spent 40 years teaching the grid to predict the weather.

And then we filled every house with a robot that argues with it.

Give me the one thing a forecasting team should do Monday morning.

Go find every rate change and demand response program your company launched in the last

18 months.

Then check whether your model has any variable that reacts to them.

If it doesn't, and it probably doesn't, your forecast isn't modeling the grid.

It's modeling a grid that stopped existing years ago.

Uncomfortable and useful.

Thanks for coming in.

Sources for today's episode.

The new stack.

MIT Technology Review.

MIT Sloan Management Review.

DBT Labs.

Vendor.

Data tooling.

Done.

Want to know where you actually stand?

Take the CDO self-assessment at MBA-training.com.

Read and go deeper