Forecasting a load that fights back: EVs, DERs, and the variance you can explain

Load forecasting used to be a tractable problem: weather, calendar, history, a smooth curve. Then the demand side started fighting back. Electric-vehicle charging piles onto the evening peak in clusters. Rooftop solar and behind-the-meter storage quietly erase midday load and hand some of it back at dusk. The curve a forecaster is chasing is being rewritten by millions of small, distributed decisions — and a model trained on the old shape degrades quietly.

Keeping up isn’t about a fancier algorithm. It’s about forecasting at a resolution, and with an honesty, the old approach never needed.

Granularity: interval and premise, not system and month

A system-level monthly forecast hides exactly the behavior that now matters. The work is to forecast at interval level and down to the individual premise, and to explicitly incorporate the new load: EV charging profiles and distributed generation, modeled as first-class inputs rather than noise the model has to absorb.

Variance you can attribute, not just measure

The real differentiator isn’t the point forecast — it’s what you do with the miss. A mature forecasting practice automates variance analysis across horizons (long-term, short-term, near-term) and between forecast and actual, and then attributes each variance: how much was weather, how much was DER, how much was EV load, how much was model error. A forecast that’s wrong and can’t say why teaches you nothing. A forecast that’s wrong and explains itself gets better every cycle — and earns the trust of the planners and traders who have to act on it.

That attribution loop is what separates a forecast people quietly override from one they plan against. It turns each miss into a correction rather than an embarrassment.

The grid edge isn’t getting simpler, and neither is the load. The utilities that stay ahead won’t be the ones with the cleverest single model — they’ll be the ones forecasting fine-grained, folding DER and EVs in on purpose, and able to explain every variance well enough to fix it.