Skip to content

Does an availability state improve the solar-farm forecast band during and after outages?

Outages of the six metered solar farms are too rare in the scored year to test an availability state, so the experiment stopped at its stop rule without fitting any model. National Grid Electricity Distribution (NGED) meters six solar farms, labelled A to F here. We asked whether a farm's state (running, partly down, or switched off), inferred from its last 6 to 48 hours of power, improves the lower quantiles (0.01 and 0.05) of the day-ahead forecast band during and after an outage. The stop rule says to stop if fewer than 200 outage half-hours fall in the scored year (2025-09 to 2026-08) at 4 or more farms. An outage half-hour is a daylight half-hour (sun more than 5 degrees above the horizon) in which the farm produced under 5% of the physical power model's forecast, when that forecast exceeded 15% of the farm's capacity, on a day with at least 6 such half-hours. No farm has 200: five have at most 34, and farm E has 165.

Results

  • Outage half-hours in the scored year (daylight, 12 to 36 h leads). A 15 (2 days), B 14 (2 days), C 0, D 34 (5 days), E 165 (15 days), F 10 (1 day). All six are under 200, so the stop rule fires.
  • A looser count gives the same answer. Counting every low daylight half-hour, without the 6-per-day requirement, gives 26 (A), 22 (B), 5 (C), 227 (D), 215 (E), and 21 (F). Farms A, B, C, and F stay under 200. Farms D and E pass 200 only because the loose count includes many short dips on 96 days (D) and 47 days (E), which the 6-per-day requirement removes.
  • Months before the scored year hold more low half-hours at some farms. The loose count before 2025-09 is 118 (A), 46 (B), 47 (C), 515 (D), 50 (E), and 78 (F).
  • Not run. The availability-state arms, their pinball loss, coverage, false alarms, and the control that shifts the power by one week.
  • The count preceded the plan commit. e00301_count.py ran before the plan was committed, so the stop rule acted as a pre-check and not as a pre-registered test.

Checks

  • Forecast checks do not apply. No forecast was made, so there is no median-minus-mean forecast and no mean-forecast error to read beside a quantile-score contrast.
  • No audit trigger fired.

Caveats and next steps

  • Label. The outage label uses hindsight output against the plant forecast. A looser label finds more outage half-hours at farms D and E, and a stricter label finds fewer.
  • Next step. If outages matter for curtailment, test the availability-state model on the full history of farms A, B, and F, scoring outage half-hours only with blocked folds. That test needs its own plan.