Surge grid

Balancing authority (BA)

A company or agency that keeps a region's grid in balance — matching electricity generation to demand in real time. PJM runs the DC-to-Chicago corridor; CAISO runs California; ERCOT runs most of Texas. 53 BAs publish a live demand feed to the US Energy Information Administration (EIA-930) — surge forecasts all of them.

Interconnection

The three giant synchronised AC grids of North America — Eastern, Western, and Texas. They barely talk to each other (only a handful of DC ties). A blackout in one can't easily spread to the others.

Day-ahead forecast

How much electricity the model thinks will be consumed over the next 24 hours, made the day before. Grid operators use day-ahead forecasts to decide which power plants to start up, how much fuel to buy, and how to price wholesale electricity.

MW and GW (megawatts, gigawatts)

Units of instantaneous power demand. 1 GW = 1 000 MW ≈ 1 large nuclear reactor at full output, or roughly enough to power 750 000 typical US homes on a hot afternoon. PJM peaks at ~165 GW; a small utility like Homestead, FL peaks at ~100 MW.

Median · p10–p90 (the shaded band)

Surge returns a probability range, not a single number. The median is the middle guess. The p10–p90 band covers the middle 80% of likely outcomes — roughly: there's a 10% chance actual demand comes in below the band, 10% above it. A narrow band means the model is confident; a wide band means weather or weekday uncertainty matters.

% of all-time peak

How close tomorrow's forecasted maximum is to that BA's historical record. 100% = likely to tie or break the peak — that's when grid stress is real. <60% is a comfortable shoulder day.

RTO / ISO

Seven US regions run competitive wholesale electricity markets (PJM, CAISO, ERCOT, MISO, NYISO, ISO-NE, SPP). Generation and retail are separate companies bidding through an auction. Non-RTO BAs are vertically integrated — one utility owns everything from the power plant to your meter.

MASE (model accuracy score)

How well the model does compared to a naive "same as last week" baseline. Lower is better. 1.0 = no better than naive; 0.5 = half the error of naive; 0.0 = perfect. surge-fm-v3 scores 0.52 on the 7 biggest grids — about 10× better than a typical weather-adjusted regression.