Offshore wind, Thu 12 Nov
1,779 MW
average over the day
Highest hour
1,974 MW
at 22:00
Fleet size used
2,036 MW
highest hour in the period before
Forecast for Thursday 12 November, made 11 Nov 10:27 (Danish time)
LightGBM on ECMWF, ICON-EU and GFS weather, made before the noon auction. Energinet's forecast is added when it is published, after the auction. How it works: Method tab.
Track record
Scored from the day after the first forecast: first with Energinet's real-time output, then with the settled numbers 8 to 10 days later.
Why only offshore
Onshore wind and solar were tried with the same rules but didn't beat Energinet on 2026. See the backtest.
How tomorrow's forecast is made, every morning
1
07:00
ECMWF's morning run is out
2
10:07
Weather at 9 points goes into the model
3
Share × fleet
The model gives output as a share of the fleet
4
Saved
Before the noon auction, never changed
5
Scored
Real-time output the next day, settled numbers 8-10 days later
The model
| Model | LightGBM (gradient-boosted trees), absolute-error loss, 1,200 trees at learning rate 0.025 |
|---|---|
| Target | Offshore output as a share of the fleet, times the fleet size |
| Weather | ECMWF's 00 UTC run from the morning before (ready about 07 UTC), plus ICON-EU and GFS. Wind at 100 m and 10 m, direction, gusts and the spread over the points |
| Points | Horns Rev, Thor and Anholt at sea, plus six points over Jutland and Funen |
| Fleet size | The highest hourly output in the 60 days before, so new parks such as Thor are picked up |
| Training | All hours from March 2024 up to 10 days before the first of the month (settled output is 8-10 days late). Refit monthly |
| Chosen on | 2025 only. Tested once on 2026, then frozen |
| Scored by | Mean absolute error per hour against Energinet's settled output; Energinet's own day-ahead forecast as the yardstick |