Forecast

What visitors see

• Next 24 hours: temperature and a weather icon per hour, with the chance of precipitation when it is 10% or more and the solar irradiance in W/m² for daylight hours (the average over the hour ending at that time). The first entry is "now".
• 7-day forecast: today and the next six days, each with an icon, the chance of precipitation and a low-to-high temperature bar.
• From this station's barometer: a one-line outlook from your own pressure readings (see The station forecast).

The Overview headline also takes its weather icon ("Partly cloudy", "Light rain") from the forecast. Units and labels follow your site's units and language.

Where the forecast comes from

The forecast is Open-Meteo's for your station's coordinates, refreshed about every 30 minutes, with a "Forecast data by Open-Meteo" note under it. It is a model prediction, so it can differ from what your station ends up recording.

By default Open-Meteo blends the best models for your location. Under Setup → Machine learning you can pick one instead: ECMWF IFS, DWD ICON, NOAA GFS, KNMI HARMONIE, DMI HARMONIE or UK Met Office. Most options show how far that model has been off at your own station. A change reaches your site within five minutes.

The local forecast (machine learning)

A weather model forecasts for a grid square, not for your garden. A station in a valley or near the coast often runs warmer or colder than the model says, and it misses in the same way at the same times of day. The Local Forecast Engine learns those habits and corrects for them.

How it works

Every hour, for each published site with coordinates, the engine blends five weather models (ECMWF IFS, DWD ICON, NOAA GFS, KNMI HARMONIE and DMI HARMONIE), weighting the ones that have done best at your station, corrects the blend with what your station measures now, and later scores both against what it recorded. Only the hourly temperature, chance of rain and solar irradiance come from the engine; daily highs and lows, rain amounts and icons stay Open-Meteo's.

Solar irradiance is the exception to the rule below: it is shown even before a correction has earned its place, because the blend itself is already weighted by how each model has done against your station's own solar sensor. Where a correction or the reading from your sensor right now has earned its place for the next hours, it is in the number too. A station without a solar sensor gets an even blend of the models.

A correction is used only where it has proved itself at your station. Each forecast range (1-6 hours ahead, 7-24, 25-48, 2-5 days and 5-10 days) earns it separately, once, within the last 30 days:

• at least 200 forecast hours in that range have been scored,
• over at least 3 days,
• and the corrected forecast was more accurate than the blend.

This is re-checked every hour, and a range that stops beating the blend loses its correction. Short ranges can qualify within days; the 5-10 day range takes longer, because the weather it predicted has to happen first.

The "Next 24 hours" strip shows "Pro Weather · Open-Meteo" when values from the engine are in it. If the engine has not produced a forecast for six hours, your site falls back to plain Open-Meteo.

Following along

Setup → Machine learning shows where your station stands:

• Solar irradiance on your site: whether the sunlight row on your Forecast tab beats Open-Meteo at your station, as its average error over the next 24 hours against the best single Open-Meteo model there. Shown once your solar sensor has scored it.
• Local forecast learning: three stages (model ready, proving itself at your station, on your Forecast tab), then for temperature, rain probability and solar irradiance, per range, your progress toward 200 hours and 3 days and the error with and without the correction, marked learning, earned or ensemble kept. Earned temperature ranges are on your Forecast tab.
• How the models score here: each model, the blend and the corrected forecast against your station over the last 7, 30 or 90 days, for temperature and, with a solar sensor, for solar irradiance in daylight hours. Lower is better.

Your data

The model learns from past forecasts and what stations then recorded. Your readings are not shown to other owners or published by this. Stations without coordinates and indoor sensors (as flagged by the source, such as WeatherLink) are left out. There is no setting to exclude a station.

The station forecast

Above the Open-Meteo forecast, "From this station's barometer" gives the Zambretti reading for your current sea-level pressure, its three-hour trend, the wind direction and the season: one of 26 phrases from "Settled fine" to "Stormy, much rain", the reading CumulusMX users know.

It is less accurate than the model forecast, but it keeps working when Open-Meteo does not and needs only a barometer: a pushed station with no coordinates still gets a Forecast tab with this line. Southern-hemisphere stations get the wind and season corrections flipped, and a three-hour change under 1 hPa counts as steady.

Turning it on or off

The forecast is on by default. Switch it under Website → Homepage → Sections and tabs → Forecast. The barometer line does not depend on this switch.

The switch needs your station's coordinates. If it is greyed out, your station has no location yet:

• WeatherLink: set the location in WeatherLink, then choose Reconnect under Setup → Connection → Cloud services.
• Ecowitt Cloud, Tempest: set it in the vendor account and reconnect.
• Pushed hardware: enter latitude and longitude under Setup → Connection → Local hardware.

The same coordinates unlock Sun & moon, Sky events and severe-weather warnings.