Firmlight

How much of tomorrow's sun can the grid count on?

Firmlight forecasts solar power across Great Britain 1 to 3 days ahead. Every forecast comes with a band, and the band was tested on a full year of real output the model never saw. Operators get one number to plan on, the power that turns up 9 hours in 10.

Tested on a year it never saw

Judging in a hurry? Three clicks.

01, live forecast

The next three days

Loading.

DayBand
Most likelyBand
Why this number?

Click any hour on the chart. Each bar shows how far one group of inputs pushed that hour's forecast up or down from the model's starting point.

The day in plain words

    02, replay

    Replay any day of a year it never saw

    Pick a day between . None of them were used in training. Compare what the model said beforehand with what the panels really produced.

    IssuedBand

    All 366 test days. Taller and darker means a bigger miss that day. Click any day to open it.

    What happenedForecastBandBaseline, same hour on the last known day

    03, proof

    Three bars, written down before training

    The pass or fail lines were fixed before the first model ran. These are the day-ahead results on the daylight hours of the test year.

    Every lead time

    Error by month, MW
    FirmlightBaseline
    How often the 80% band held
    The dashed line is the 80% target.
    Is each band honest?
    What drives the forecast

    How much worse the day-ahead forecast gets, in MW, when one group of inputs is scrambled on the test year.

    Under every kind of sky

    Day-ahead test hours split into thirds by how clear the forecast sky was.

    04, planner

    Turn the forecast into a reserve decision

    Choose how often you can accept solar coming in under plan. Firmlight commits the bottom of the matching band and holds the gap as reserve. The baseline is held to the same shortfall rate, so the comparison is fair. Every figure is recomputed in your browser from the test year.

    Shortfall you acceptLead time
    An example number. Change it to yours.
    Combined cycle gas, from the NESO carbon intensity method. Change it if you have a better figure.

    05, storage

    How big a battery makes the promise firm?

    Some hours still come in under the firm line. A battery can cover them. This sizes that battery from every shortfall in the test year, for Firmlight's firm line and for the baseline's, at the risk and lead time chosen in the planner above.

    Days the battery must cover
    FirmlightBaselineEach test day's shortfall energy, sorted from the worst day to the best.

    06, regions

    Fourteen regions, one model

    One pooled model forecasts each of the 14 grid regions a day ahead. Pick a region. Your browser computes its forecast for tomorrow. The second chart replays the day you chose in the replay section, for that region.

    Loading the regions.

    Tomorrow

    Replay day

    What happenedForecast80% bandBaseline

    07, method and limits

    How it works, and where it falls short

    • Generation. Hourly national solar output and installed capacity from Sheffield Solar PV_Live, Jan 2024 to Nov 2026.
    • Weather. Open-Meteo Previous Runs archive at 8 points across England. These are the forecasts as they were issued 1, 2 and 3 days earlier, so each lead time is scored on what was knowable then.
    • Model. Gradient boosted quantile regression on capacity factor, 7 quantiles per lead time. Inputs are forecast sunlight, cloud and temperature, the spread between sites, sun position, and the older forecasts of the same hour.
    • Calibration. Conformalized quantile regression. Every fifth day before the test year is held out to widen each band until it holds.
    • Model check. Loading the browser copy of the model.