Forecasts and models¶
A station or point page shows five weather models side by side and opens with the one that was closest to the measurements there. This page explains where the numbers come from and how "closest" is decided.
The models¶
| Model | By | Resolution | Ahead | Runs a day | Class |
|---|---|---|---|---|---|
| ICON-D2 | DWD | 2.2 km | 48 h | 8 | Superforecast (Germany and the Alps) |
| AROME | Météo-France | 1.5 km | 42 h | 8 | Superforecast (France and around) |
| HARMONIE | DMI | 2 km | 60 h | 8 | Superforecast (Europe) |
| ECMWF IFS | ECMWF | 25 km | 15 days | 4 | Forecast (world) |
| GFS | NOAA | 13 km | 16 days | 4 | Forecast (world) |
A regional model shows only where its domain covers the place. WF_MODELS limits the list, or
the switches on the admin page when the variable is not set.
How forecasts are fetched¶
Requests go to Open-Meteo only from the server, never from the browser. A place (a station, or a point rounded to the 0.01° grid) is fetched when someone opens it, once per model run: Open-Meteo's model metadata says when a new run is out, and a place is asked again only then. Places opened within the last seven days are refreshed after every new run; points nobody opened for two days are dropped. A model that answers a place without values (a regional model's rectangle reaches past its domain) is asked there again a day later, not with every run.
The free API allows about 10,000 calls a day (WF_OPEN_METEO_DAILY_BUDGET, 9,000 by default):
enough for a few hundred places people look at regularly. Opening new places may use two thirds
of it, at most 30 an hour and 100 a day per client; the rest stays for the refreshes
(limits). The calls are counted per UTC day in
the database, so a restart does not reset them; the admin page shows the day and the last 14, and
changes the seven and two days, the budget and the limits without a restart.
windfenster probe --lat … --lon … prints what Open-Meteo returns for every model at a place.
The ten days¶
The page builds the ten days from the selected model and fills what it does not reach from the global one:
- Today and tomorrow: hourly from 7 to 19 when a fine regional model is selected (ICON-D2, AROME, HARMONIE), every 3 hours with a global one or the mix.
- Days 3 to 7: every 3 hours around the clock, 8 columns.
- Days 8 to 10: the same columns, marked Trend.
The quick view takes the same days every 3 hours: the bars, the main direction over the daytime (the mean of the wind vectors, so a strong hour counts more than a calm one and north stays north across 360°), and the strongest daytime wind.
Verification and ranking¶
Every hour Windfenster compares, for each station with measurements, every stored forecast hour with the measurement closest to it (within 20 minutes):
- the mean wind: absolute error and bias in kn;
- the direction: the angle between them, from 5 kn measured on;
- "14 kn or more": whether both said so.
A run counts only for the hours after Windfenster fetched it, so a model that arrives late gains nothing from hours that had passed. Errors are kept per station, model and UTC day, split by lead time: 0–24, 24–48 and 48–72 hours after the run. The TAF counts as a model of its own, from its issue time.
flowchart LR
F[stored forecasts<br>per run] --> P{pair each hour<br>with the nearest<br>measurement ±20 min}
O[measurements] --> P
P --> E[errors per station,<br>model, day, lead time]
E --> R[ranking:<br>MAE of the first day,<br>last 30 days]
E --> A[auto choice]
E --> W[mix weights<br>1/MAE²]- Who hits here?: the mean absolute error of the first day ahead over the last 30 days. A point takes the ranking of a station at most 5 km away.
- Auto choice: from 14 days of data the page opens with the regional model of the lowest error for 0–24 h, and Forecast takes the global model of the lowest error for 48–72 h. Before that a fixed order decides (ICON-D2, AROME, HARMONIE, then ECMWF IFS, GFS). A saved spot can name its own starting model instead (the spot's dialog, Starting model); where that model has no values, the auto choice holds.
The Windfenster mix¶
Per hour the weighted mean of every model with a value there, each weighted by 1/MAE² of that model at the station for that lead time (all alike before 14 days); the direction is that of the mean wind vector. The mix is computed when a page is served and verified like a model, as it stood each time new runs came in. It opens the page only where it beat the best single model.
Do the models agree?¶
The spread (highest minus lowest mean wind) of the models at the windiest daytime hour:
| Spread | Agreement |
|---|---|
| up to 4 kn | high |
| up to 8 kn | medium |
| more | low |
The tile shows it for the next two days with every model on a strip; each day card and each quick-view tile show it for their day.