Weather products surveyed for the weather-product studies

This page is a survey, not a plan: it records which weather products could join two studies of past weather and a planned study of forecasts (issue #810), and commits the project to ingesting none of them. The list of products the project ingests, plans to ingest, or is researching as a possible source is the catalogue under Weather data.

What the project uses

The live forecast uses one weather product, ECMWF ENS, and two more are planned. The studies score 13 products in total: the seven products in this table, four more Open-Meteo models from the survey below, ICON-DREAM-EU from the reanalysis table below, and SARAH-3, which the catalogue lists and this page does not. Every other product on this page is surveyed only, except where the "Scored in" column says otherwise. "Roadmap status" copies the product's row in the Weather data catalogue; a version label such as v0.5 names a roadmap milestone.

Product	What it is	Roadmap status	What the project uses it for	Archive read from	Scored in
ECMWF ENS	The European Centre for Medium-Range Weather Forecasts' (ECMWF's) 51-member ensemble forecast	✅ Ingested today	The live forecast	Dynamical.org, a non-profit that republishes weather-model archives	Sunshine
ERA5	ECMWF's reanalysis: a weather model re-run over the past and pulled towards observations	🚧 Planned ingestion, milestone v0.5	Planned: training history and capacity estimation (why ERA5 scores worse than current products)	Open-Meteo's copy, for the studies; the planned ingest route is still open	Sunshine, wind
CAMS	The Copernicus Atmosphere Monitoring Service's satellite irradiance	🚧 Planned ingestion, milestone v0.7	Planned: offline capacity estimation only (which CAMS route, and its traps)	The CAMS Radiation Service point API	Sunshine
ICON-EU	The German weather service's (DWD's) ICON model, 6.5 km Europe configuration	🔬 Research (v0.9, uncertain), as Dynamical.org's copy	Studies only	Open-Meteo's Historical Forecast archive	Sunshine, wind
UKV	The Met Office's 2 km UK model	🔬 Research (uncertain), as the Met Office's AWS feed	Studies only	Open-Meteo's Historical Forecast archive	Sunshine, wind
ICON-D2	DWD's ICON model, 2 km central-Europe configuration	Not on the roadmap	Studies only	Open-Meteo's Historical Forecast archive	Sunshine, wind
ICON global	DWD's ICON model, global configuration	Not on the roadmap	Studies only	Open-Meteo's Historical Forecast archive	Sunshine, wind

CAMS describes past sunshine best, and UKV and ICON-D2 describe past wind best. Which weather product best describes past sunshine? scores eight products from December 2022 to August 2026 and finds CAMS best by a wide margin, and ICON-D2 the best of the weather models as served. The page also scores four more Open-Meteo models on a shorter window, from November 2024 to August 2026. Which weather product best describes past wind? scores August 2024 to September 2026 and finds UKV and ICON-D2 best. CAMS publishes no wind, so the wind study scores five of the products in the table above, and a sixth, ICON-DREAM-EU, on a slightly shorter row set.

Terms used on this page

The survey's comparisons turn on how a forecast is run and archived, so these terms are defined before any section uses them:

Run — one execution of a weather model from a start time, its initialisation, producing values for the hours ahead: "the 00 UTC run".
Lead time — how far ahead of its run's start a value lies. "T+54 h" is the value 54 hours after the run started.
Fixed lead — every product compared at the same lead time, so a difference between products is not a difference in how old each forecast was.
Whole run — every lead time of a run, kept in an archive, as opposed to only the freshest run's value for each hour.
Hindcast — a forecast re-run after the fact with a later model version.
Serving path — the code that produces the live forecast each day, as opposed to training and offline analysis.
Global, direct, and diffuse irradiance — global irradiance is all the sunlight reaching a horizontal surface. The direct beam is the part arriving straight from the sun, and diffuse is the rest, scattered by cloud and air. A solar-farm model needs the split, because tilted panels catch the two parts differently. Direct normal irradiance (DNI) is the direct beam measured facing the sun.
Period-ending — how the meters label each half-hourly reading: with the time the half-hour ends. A product stamped at the middle of its averaging period needs shifting to match.
HRES — ECMWF's single 9 km forecast from its Integrated Forecasting System (IFS). Since IFS Cycle 50r1 went live on 2026-05-12, ECMWF no longer runs a separate HRES: the single forecast is the ENS control member (ECMWF, plans). Open-Meteo still calls the product "IFS HRES 9 km", and so does this page. AIFS is ECMWF's machine-learned Artificial Intelligence Forecasting System.

How to read the Open-Meteo rows

Open-Meteo, the service the studies read, re-serves other centres' forecasts, so a start date, a latency, or a direct/diffuse split read through Open-Meteo may belong to the archive rather than to the producer. The tables below say which. The production service does not read Open-Meteo. Four conventions apply to every Open-Meteo row:

Start dates are the first hour with a non-null value at Lincoln (53.23°N, 0.54°W), queried on 2026-09-23. Lincoln stands in for the trial area in the East Midlands, so a start date, a null series, or a ratio measured there is a result for that one point. A user elsewhere in Great Britain should re-query at their own sites, and the "Covers GB" column says where coverage was checked on the grid rather than at Lincoln. ICON-D2, for example, has data at Lincoln but none in the west of Great Britain: the map of ICON-D2's and AROME France's domains on the data sources page shows where.
Open-Meteo's UKV archive is a backfill before 12 August 2024, so a UKV start date on this page is not the start of UKV's own run history (what the backfill is and how it changes scores).
Latency is the time from a run's initialisation to the whole run being available on Open-Meteo, read from that model's meta.json for its latest run on the morning of 2026-09-23: 00 UTC for the 6-hourly models, a later run for AROME, the Danish Meteorological Institute (DMI), and the Royal Netherlands Meteorological Institute (KNMI). Each figure is one run, not an average.
"Native" means the producer publishes the direct or the diffuse field, and "separation" means Open-Meteo derives both from global irradiance with the Razo–Müller–Witwer separation model (Open-Meteo ECMWF documentation). A split made by separation carries no information the global field lacks.
The Licence column gives the licence of the data. Open-Meteo's free API is for non-commercial use only (terms), so commercial use needs a paid Open-Meteo plan.

Which archives keep whole past forecast runs

Issue #810 compares models at a fixed lead, so it needs an archive that keeps whole past runs or serves a fixed lead. A model run every 3 hours and a model run every 6 hours reach a given hour at different ages, so a series that always takes the freshest run compares delivery cadence as well as forecast quality. Open-Meteo publishes three archives of the same forecasts, and Dynamical.org publishes a fourth kind:

The Historical Forecast archive stitches the first hours of each successive run into one continuous series, so a product's lead follows its run frequency rather than being chosen (Historical Forecast API). The two study pages score this archive.
The Previous Runs archive serves each variable at a fixed offset of 1 to 7 days (Previous Runs API). The archive starts on 2024-01-19 at the earliest, and serves day 1 only for UKV, ICON-D2, the Met Office global 10 km model, and the HARMONIE-AROME and AROME models (table below).
The Single Runs archive keeps whole runs (Single Runs API), but only from the 2026-04-02 00 UTC run: ecmwf_ifs025 and ncep_gfs013 both reject run=2026-04-01T00:00. ECMWF IFS HRES 9 km is the exception, below.
Dynamical.org keeps whole runs as Icechunk (Zarr) repositories, listed in its SpatioTemporal Asset Catalog (STAC). The start dates below come from Dynamical.org's catalogue pages, checked against each store's init_time coordinate where the date matters here. The longest-running Dynamical.org stores reach back to 2020-10-01 for GEFS, the US National Oceanic and Atmospheric Administration's (NOAA's) Global Ensemble Forecast System, and to 2021-05-01 for GFS, NOAA's Global Forecast System.

Open-Meteo's Single Runs archive keeps ECMWF IFS HRES 9 km from 2024-03-14, two years earlier than the other models. The model ecmwf_ifs rejects run=2024-03-13T00:00 and serves the 2024-03-14 00 UTC run with 240 hourly steps. Open-Meteo calls the runs from 2024-03-14 "IFS Cycle 49R1 hindcasts". The March to June 2024 runs sampled are 00 and 12 UTC only, and 06 UTC runs are present from 2024-08-10. Runs from 2026-05-12 06 UTC use IFS Cycle 50r1 (what that upgrade changed).

Dynamical.org stops serving data.dynamical.org on 2026-09-30, so any URL on that host has to move to a supported access pattern. Every Dynamical.org catalogue page carries the notice "data.dynamical.org access ends September 30, 2026".

Which offsets the Previous Runs archive serves is set by the archive, not by the model's horizon. UKV's runs in the Single Runs archive hold 55 hourly steps, to T+54 h, and ICON global's runs reach 180 h, yet the Previous Runs archive serves UKV at day 1 only and ICON global to day 6 only. Any lead-matched comparison is therefore capped at day 1 where one arm is UKV, ICON-D2, the Met Office global 10 km model, or a HARMONIE-AROME or AROME model.

Model on Open-Meteo	First day-1 global irradiance at Lincoln	Offsets served
ICON-D2	2024-01-19	Day 1
Météo-France AROME France	2024-01-19	Day 1
KNMI HARMONIE-AROME Europe	2024-06-27	Day 1
DMI HARMONIE-AROME DINI	2024-06-29	Day 1
UKV	2024-08-06	Day 1
Met Office global 10 km	2025-01-03	Day 1
Météo-France ARPEGE (Europe and World)	2024-01-19	Days 1–3
ICON-EU	2024-01-19	Days 1–4
ICON global	2024-01-19	Days 1–6
NOAA GFS	2024-01-19	Days 1–7
ECMWF IFS 0.25°	2024-03-06	Days 1–7
ECMWF AIFS Single	2025-02-17	Days 1–7
ECMWF IFS HRES 9 km	2025-10-01	Days 1–7

Whole-run archive	Runs kept	First run	Radiation and wind in the archive
NOAA GEFS 35-day (Dynamical.org)	31 members, 00 UTC only, to 840 h	2020-10-01	Global irradiance; 10, 80, and 100 m wind
NOAA GFS (Dynamical.org)	4 a day, to 384 h	2021-05-01	Global irradiance; 10, 80, and 100 m wind
Google WeatherNext 2 (Earth Engine)	64 members, 4 a day, to 15 days	2022-01-01	No radiation; 10 and 100 m wind
ECMWF IFS HRES 9 km (Open-Meteo Single Runs)	00 and 12 UTC at first; 06 UTC from 2024-08-10	2024-03-14	Global and native direct irradiance; 10, 80, and 100 m wind
ECMWF IFS ENS (Dynamical.org)	51 members, 00 UTC only, to 360 h (906 runs to 2026-09-23)	2024-04-01	Global irradiance; 10 and 100 m wind
ECMWF AIFS Single (Dynamical.org)	4 a day, to 360 h	2024-04-01; radiation and 100 m wind from the 2025-02-24 06 UTC run	Global irradiance; 10 and 100 m wind
ECMWF AIFS-ENS (Dynamical.org)	4 a day, to 360 h	2025-07-02	Global irradiance; 10 and 100 m wind
DWD ICON-EU (Dynamical.org)	00, 06, 12, and 18 UTC, to 120 h	2026-02-10	Native direct and diffuse irradiance; 10 m wind only
CEDA NWP-UKV, the Centre for Environmental Data Analysis	Up to 8 a day, T+0 to T+120	2016-03-16	Radiation not listed on the catalogue record; access by application

The surveyed products

None of the products from here on is used by the project, except where its Roadmap status says so. The tables give each product's roadmap status in words beside the catalogue's symbol, and the study pages that score it. "Grid" is the producer's native grid spacing, with the serving grid in brackets where the two differ. "Past runs" names the archive that keeps whole runs or a fixed lead, from the section above.

Of the forecast models in the tables below whose solar variables were established, only ECMWF's IFS 9 km forecasts, GFS, the Met Office global model, ICON-EU, and Google WeatherNext 3 serve a producer's own direct or diffuse field. For the other models Open-Meteo serves, Open-Meteo makes the split by separation, except for DMI HARMONIE-AROME, whose served direct field is probably not a direct beam at all (note below the first table). The Dynamical.org and Google products serve global irradiance only, or no radiation. UKV, ICON-D2, and ICON global, which the studies score, also serve the producer's own direct field, as do UKV and MOGREPS-UK on the Met Office's own feed — see which sources carry which irradiance components.

Among the reanalyses surveyed, CERRA and COSMO-R6G2 carry the producer's own direct field, as ERA5 and ICON-DREAM-EU, both already scored, do.

Forecast models

Product	Roadmap status	Scored in	Producer and method	Grid	Covers GB	Archive start	Solar variables	Wind heights	Runs, horizon, step	Latency	Past runs	Access	Licence	Source
ECMWF IFS ENS (Dynamical.org)	✅ Ingested today	Sunshine only	ECMWF, 51-member physics ensemble	9 km (0.25° served)	Yes (global)	2024-04-01 to present; back-fill in progress (catalogue row)	Global only (`ssrd`)	10, 100 m	4 a day from ECMWF, Dynamical.org keeps 00 UTC; 360 h; 3-hourly to 144 h, then 6-hourly	Not established	Whole runs	Icechunk (Zarr)	CC BY 4.0 and ECMWF terms	Dynamical.org
ECMWF IFS ENS 9 km Europe (Open-Meteo, `ecmwf_ifs_europe_ensemble`)	Not on the roadmap (the catalogue's ENS row is the Dynamical.org feed)	Not scored	ECMWF, 51-member physics ensemble, European domain	9 km, O1280 (served natively)	Data at Lincoln	2026-09-23 01 UTC at Lincoln, asking for 7 past days on 2026-09-23	Global; direct native (ECMWF's `fdir`, per Open-Meteo's downloader)	Not checked	00 and 06 UTC; 15 days; hourly to 90 h, 3-hourly to 144 h, then 6-hourly	Not established	None beyond the latest runs	Open-Meteo Ensemble API	CC BY 4.0 (ECMWF's licence and delivery terms)	Open-Meteo, downloader
ECMWF IFS HRES 9 km (`ecmwf_ifs`)	Not on the roadmap	sunshine only, on a shorter window	ECMWF, global physics model	9 km (served natively)	Yes (global)	2017-01-01, complete to 2026-09-21	Global; direct native (ECMWF's `fdir`, per Open-Meteo's downloader); diffuse = global − direct	10, 80, 100 m	00 and 12 UTC to 15 days, 06 and 18 UTC to 144 h; hourly to 90 h, 3-hourly to 144 h, then 6-hourly	7.0 h (Open-Meteo)	Single Runs from 2024-03-14; Previous Runs from 2025-10-01	Open-Meteo	CC BY 4.0 (ECMWF's licence and delivery terms)	Open-Meteo, downloader
ECMWF IFS 0.25° (`ecmwf_ifs025`)	Not on the roadmap	Not scored	ECMWF open-data subset of the same run	9 km (0.25° served)	Yes (global)	Radiation and 100 m wind 2024-03-05; 10 m wind 2024-02-03	Global; direct and diffuse by separation	10, 100 m	00 and 12 UTC to 15 days, 06 and 18 UTC to 144 h; 3-hourly to 144 h, then 6-hourly	7.8 h (Open-Meteo)	Previous Runs from 2024-03-06; Single Runs from 2026-04-02	Open-Meteo	CC BY 4.0	Open-Meteo
ECMWF AIFS Single	Not on the roadmap (the catalogue's AIFS-ENS row is the ensemble)	Not scored	ECMWF, machine-learned deterministic model; v1 operational from the 2025-02-25 06 UTC run, adding `ssrd` and 100 m wind	N320, about 0.25° (0.25° served)	Yes (global)	Open-Meteo 2025-02-17; Dynamical.org radiation from 2025-02-24 06 UTC	Global only (`ssrd`); Open-Meteo's direct is by separation	10, 100 m	4 a day; 360 h; 6-hourly	5.7 h (Open-Meteo)	Whole runs (Dynamical.org); Previous Runs from 2025-02-17	Open-Meteo; Icechunk (Zarr)	CC BY 4.0 and ECMWF terms	Dynamical.org, ECMWF
ECMWF AIFS-ENS	🔬 Research (v2.1, uncertain)	Not scored	ECMWF, machine-learned ensemble	0.25°	Yes (global)	2025-07-02	Global only (`ssrd`)	10, 100 m	4 a day; 360 h; 6-hourly	Not established	Whole runs	Icechunk (Zarr)	CC BY 4.0 and ECMWF terms	Dynamical.org
DWD ICON-EU (Dynamical.org)	🔬 Research (v0.9, uncertain)	Both studies, through Open-Meteo's copy (sunshine, wind)	DWD, regional physics model	6.5 km (0.0625° served)	Yes	2026-02-10	Direct and diffuse native	10 m only in Dynamical.org's store; Open-Meteo serves 80 and 120 m	Every 3 hours on Open-Meteo, Dynamical.org keeps 00, 06, 12, and 18 UTC; 120 h; hourly to 78 h	Not established	Whole runs (Dynamical.org); Previous Runs from 2024-01-19 (Open-Meteo)	Icechunk (Zarr); Open-Meteo	CC BY 4.0	Dynamical.org, Open-Meteo
NOAA GFS (`ncep_gfs013` for surface fields, `ncep_gfs025` for 80 and 100 m wind)	Not on the roadmap	Not scored	The National Centers for Environmental Prediction of the US National Oceanic and Atmospheric Administration (NOAA NCEP), global physics model	~13 km (0.11° served for surface fields, 0.25° for 80 and 100 m wind)	Yes (global)	2021-03-23	Global; diffuse from NCEP's visible-band diffuse flux (`VDDSF`), not a broadband field; direct = global − that diffuse	10, 80, 100 m	4 a day; 16 days	5.5 h at 0.11°, 6.5 h at 0.25° (Open-Meteo)	Previous Runs from 2024-01-19; whole runs on Dynamical.org from 2021-05-01	Open-Meteo; Icechunk (Zarr)	Public domain from NOAA; CC BY 4.0 as served by Dynamical.org and Open-Meteo	NOAA EMC, Open-Meteo, downloader, Dynamical.org
NOAA GEFS 35-day (Dynamical.org)	Not on the roadmap	Not scored	NOAA NCEP, 31-member ensemble	0.25° to 240 h, 0.5° after	Yes (global)	2020-10-01	Global only	10, 80, 100 m	00 UTC; 840 h; 3-hourly to 240 h, then 6-hourly	Not established	Whole runs	Icechunk (Zarr)	Public domain from NOAA; CC BY 4.0 on Dynamical.org	Dynamical.org
Météo-France ARPEGE Europe	Not on the roadmap	sunshine only, on a shorter window	Météo-France, stretched-grid global model	~5 km over Europe (0.1° served)	Bounds of the Europe extract not checked; data present at Lincoln	Global irradiance and 10 m wind 2022-11-09; 100 m wind 2023-06-05	Global; direct and diffuse by separation	10, 80, 100 m	4 a day; 4 days	3.5 h (Open-Meteo)	Previous Runs from 2024-01-19	Open-Meteo	Météo-France's own licence, per Open-Meteo; CC BY 4.0 on Open-Meteo	Météo-France, Open-Meteo, licences
Météo-France ARPEGE World	Not on the roadmap	Not scored	The same ARPEGE run, global extract	~5 km over Europe (0.25° served)	Yes (global)	2023-12-27	Global; direct and diffuse by separation	10, 100 m	4 a day; 4 days	4.0 h (Open-Meteo)	Previous Runs from 2024-01-19	Open-Meteo	As ARPEGE Europe	Open-Meteo
Météo-France AROME France	Not on the roadmap	Not scored	Météo-France, convection-permitting regional model	1.3 km (0.025° served)	South of 55.4°N only	2023-12-27	Global; direct and diffuse by separation	10, 80, 100 m	8 a day; 2 days	2.7 h (Open-Meteo, 03 UTC run)	Previous Runs from 2024-01-19	Open-Meteo	As ARPEGE Europe	Open-Meteo, domain
DMI HARMONIE-AROME DINI	Not on the roadmap	sunshine only, on a shorter window	The Danish Meteorological Institute (DMI) for the United Weather Centres West (UWC-West) consortium; HARMONIE-AROME is the convection-permitting regional model several European weather services share	2 km (Open-Meteo's figure)	Yes (39.7–62.7°N, 25.4°W–40.1°E)	2024-06-28	Global; a field Open-Meteo serves as direct, probably not DMI's direct beam (note below)	10, 50, 100, 150, 250, 350, 450 m (Open-Meteo interpolates 80 m)	8 a day; 2.5 days	2.8 h (Open-Meteo, 06 UTC run)	Previous Runs from 2024-06-29	Open-Meteo	CC BY 4.0	Open-Meteo, domain, downloader
KNMI HARMONIE-AROME Europe	Not on the roadmap	sunshine only, on a shorter window	The Royal Netherlands Meteorological Institute (KNMI) for UWC-West	5.5 km, as KNMI publishes the grid	Yes (39.7–62.6°N, 25.2°W–38.8°E)	Global irradiance 2024-06-26; 100 m wind 2024-10-09	Global; direct and diffuse by separation	10 to 300 m, including 80 and 100 m	Hourly; 2.5 days	2.6 h (Open-Meteo, 06 UTC run)	Previous Runs from 2024-06-27	Open-Meteo	CC BY 4.0	Open-Meteo, domain
Met Office global 10 km (`ukmo_global_deterministic_10km`)	Not on the roadmap (why no request has been made)	Not scored	Met Office Unified Model, global	~10 km (0.09° served)	Yes (global)	Direct and 10 m wind 2022-03-01; global irradiance from 2025-01-02	Global; direct native; diffuse = global − direct	10 m only	4 a day; 7 days (Open-Meteo)	8.4 h (Open-Meteo)	Previous Runs from 2025-01-03	Open-Meteo; AWS	CC BY-SA 4.0 (AWS)	Open-Meteo
Google WeatherNext 3	🔬 Research (after v2, unlikely)	Not scored	Google DeepMind, machine-learned 64-member ensemble	0.1°	Yes (global)	2026-01-01, no back-fill before it (confirmed against the bucket 2026-09-28)	Global; direct native (`fdir`); diffuse = global − direct	10, 100 m	4 a day to 15 days, plus hourly interim runs to 48 h; hourly	Not established	Not established	Access request; members only in the Cloud Storage Zarr store	CC BY 4.0 once at least 1 h old; real-time data under Google DeepMind's experimental terms	model specs, access
WeatherNext 3 statistics (ensemble mean)	🔬 Research (study input; production unlikely before v2)	Matched-lead study, once the fetched archive is validated	Google DeepMind, the same 64-member ensemble; OCF fetches the precomputed ensemble-mean statistic, not the members	0.1° (about 11 km)	Yes (OCF's own crop, 49.0–61.5°N, 10.0°W–3.5°E)	2026-01-01, no back-fill before it (confirmed against the bucket 2026-09-28)	Global; direct native (`total_sky_direct_solar_radiation_at_surface_1hr_mean`), not by subtraction	10, 100 m	4 a day (00, 06, 12, 18 UTC); 360 h; hourly throughout, no step widening	Not established	Whole runs, in OCF's own Icechunk store	Icechunk (Zarr), `ocf-weathernext3-uk` (OCF's crop); source in Google's `weathernext3_statistics_spatial` bucket, no Requester Pays	CC BY 4.0 once at least 1 hour old, plus DeepMind's access terms	Google's GCS guide
Google WeatherNext 2	Not on the roadmap (the catalogue's Google row is WeatherNext 3)	Not scored	Google DeepMind, machine-learned 64-member ensemble; Google marks the data experimental, "not intended, validated, or approved for real world use"	0.25°	Yes (global)	2022-01-01 to 2026-09-22 in the Earth Engine catalogue	None	10, 100 m	4 a day; 15 days; 6-hourly	About 7.5 h (Google's dissemination schedule)	Whole runs	Earth Engine, BigQuery, Cloud Storage Zarr; all need an access request	CC BY 4.0 for data over 48 h old; real-time data under Google DeepMind's experimental terms	Earth Engine
CEDA NWP-UKV (the UK's Centre for Environmental Data Analysis)	Not on the roadmap (the catalogue's UKV row is the Met Office's AWS feed: 🔬 Research (uncertain))	Not scored (the studies score Open-Meteo's UKV)	Met Office UKV as archived at CEDA; statistically different from the UKV served live on AWS and Open-Meteo, so not interchangeable with it for training and inference	~0.018° (CEDA's figure)	Yes	2016-03-16 to present	Not listed on the catalogue record	Surface and pressure levels	Up to 8 a day; 120 h	Archive, ongoing	Whole runs	CEDA, by application	Met Office licence via CEDA	CEDA
CEDA Met Office global (`global-grib`)	Not on the roadmap	Not scored	Met Office global model, raw output	Not established	Yes (global)	Directories from 2016-03 to 2026-09	Not established; the README needs a login	Not established	Not established	Archive	Not established	CEDA, by application	Not established	CEDA

"Latency" in this table is the time from a run's start to the whole run being available on the named service. MARS is ECMWF's Meteorological Archival and Retrieval System.

The low direct share Open-Meteo serves for DMI HARMONIE-AROME probably comes from Open-Meteo's field mapping rather than from DMI's model. Direct over global irradiance at Lincoln was 0.24, 0.30, and 0.18 in August 2024, June 2025, and June 2026, against 0.46 to 0.62 for KNMI HARMONIE-AROME, UKV, and ECMWF IFS HRES in the same months. Open-Meteo's downloader serves DMI's "downward short-wave radiation flux" as direct and leaves DMI's "direct solar exposure" field unused, with a comment that the "direct solar exposure" field "seems to be DNI". The mapping has not been checked against DMI's own files.

What we learnt about WeatherNext's precomputed statistics store

The statistics bucket has no Requester Pays billing, unlike the assumption that first justified fetching it from a Compute Engine VM. Checking the bucket's own metadata directly, rather than inferring it from the access-request process, showed no billing field at all — egress from weathernext3_statistics_spatial is billed to Google's project, not to whichever project reads it. The VM stayed the right choice regardless: the study needs the whole 2026-to-present archive at four runs a day, and that volume is too much for a domestic internet connection irrespective of who pays for it, while a Compute Engine machine in the same region as the bucket reads it far faster.

The archive carries no back-fill before 2026-01-01. The bucket's zarr/ prefix holds only 2026_to_present, with no earlier date-range prefix, confirmed by listing the bucket directly on 2026-09-28.

Every lead time in the statistics store is hourly, with no step widening. The four daily runs (00, 06, 12, and 18 UTC) hold 360 hourly lead times each, unlike ECMWF ENS and the AIFS models on this page, which widen from 3-hourly or 6-hourly steps partway through the horizon. Both direct and total surface solar radiation arrive as named fields — total_sky_direct_solar_radiation_at_surface_1hr_mean and surface_solar_radiation_downwards_1hr_mean — accumulated in J m⁻² over the hour ending at the valid time, so dividing by 3600 gives the hour's mean irradiance in W m⁻².

OCF holds a cropped copy in its own Icechunk store, because cropping does not reduce what the source bucket transfers. Each source chunk covers one lead time and one variable over the whole globe, so pulling any part of the United Kingdom box still reads the full chunk — about 50 GB per run, and about 53 TB for the whole archive at four runs a day. The fetch crops each run down to 49.0–61.5°N, 10.0°W–3.5°E after reading it, and writes the crop into an Icechunk repository at ocf-weathernext3-uk, in the same us-east1 region as the source. In the workstation tests that preceded the VM build, a single run's read cost about £4.50 from outside us-east1 and took about 21 minutes; the same run, read from inside us-east1 on the VM, took about 2 minutes 47 seconds. Reading the whole cropped store from outside Google Cloud is a one-off egress cost of a few pounds, borne by the bucket's own project rather than split out per reader.

Each run lands on a staging branch and only reaches main once validated, so colleagues reading main see a store that has passed structural checks, per-run physical-range and radiation sanity checks, and a spot comparison against the source bucket by value.

The fetch holds a steady rate of about 29 to 31 runs an hour, with no retries or errors.

The Compute Engine VM's cost for the full archive is estimated at about £10–11, from published list pricing rather than a real invoice, because the project's Cloud Billing API is not enabled. Enabling that API for exact cost tracking, and deleting an unused BigQuery linked dataset left over from a route the project considered and did not take, were both left undone, since neither was judged worth the effort for a single archive build. The output bucket does not have Requester Pays enabled, matching the source buckets, because OCF does not split cloud costs by project internally.

AIFS has not been shown to improve faster than the physics-based IFS

In the sources we read, we found no like-for-like evidence that AIFS improves faster than the physics-based IFS. The claim would need AIFS version-to-version scores set beside IFS cycle-to-cycle scores, computed the same way against the same truth. We found no version-to-version or cycle-to-cycle scores for 2 m temperature, wind, or radiation, from either model. The search covered ECMWF's implementation pages, newsletters, and blog, one independent industry analysis, and three journal papers or preprints, and searches of arXiv and OpenAlex for independent verification. It did not cover the wider peer-reviewed literature.

ECMWF's own reports show AIFS ahead of the IFS at one point in time, not improving faster. In Forecast performance 2025, ECMWF reports AIFS error reductions against the IFS that are "typically of the order of 5–15%" in the medium range, with AIFS leading for 2 m temperature and 10 m wind speed, especially over Europe. The same page reports that ECMWF's Technical Advisory Committee found AIFS skill decreased slightly over the preceding 12 months, as it did for several other machine-learned forecasts. For AIFS ENS v1, ECMWF reports gains of up to 25% for upper-air variables and better 2 m temperature at all lead times against surface station observations, with degradations beyond day 7 in some regions. The same page reports 10 m wind and 100 hPa temperature as worse than the physics-based ensemble. Solcast, scoring against satellite-derived irradiance in the weeks after AIFS Single v1 went operational, called its own dataset limited. Solcast found a bias of about -8% for AIFS against about +2% for the IFS, and about equal accuracy at day 1, 3, and 5. Every source in this paragraph except Solcast is ECMWF's own.

The journal paper on AIFS Single 1.1.0 shows a one-off margin over the IFS in 2023, and a gain over the pre-operational AIFS, neither of which is a rate comparison. Moldovan et al. (2026) report that AIFS gains 12 to 24 hours of lead time at days 3 to 10 over the operational IFS (cycles 47r3 and 48r1), scored on Northern Hemisphere 500 hPa geopotential and 850 hPa temperature anomaly correlation against the IFS analysis for all of 2023. Moldovan et al. also report that the new AIFS improves on the pre-operational AIFS by around 4 to 6%, and that precipitation improves by up to 12% in the short range and gains about 1 day on the stable equitable error in probability space score over both the IFS and the pre-operational AIFS, against 24-hour surface station reports. Against the IFS, surface shortwave downwards radiation gains about 1 day in the medium range, scored against CM SAF geostationary satellite data. The paper reports 2 m temperature and 10 m wind as "similarly improved" against surface station observations. The truth data differ between scores, and the paper gives no year-on-year IFS cycle comparison.

Two independent preprints score AIFS against surface observations, and neither shows AIFS improving faster. Kocsis and Baran (2026) scored raw 10 m wind speed from AIFS ENS v1 against raw IFS ensemble forecasts (cycle 49r1) at 9,246 surface stations from 2025-07-01 to 2025-11-30, using 12 UTC runs. The IFS ensemble scored better at every horizon, and the gap narrowed after post-processing. Trotta et al. (2025) scored AIFS Single (the paper states no version) against the ECMWF HRES and ENS forecasts at Australian surface stations from 2024-03-01 to 2024-07-23 (569 stations, with scores from 2024-04-01), before AIFS Single was operational. Trotta et al. report that AIFS falls behind ENS as lead time grows. Apart from Solcast's, we found no independent verification of AIFS radiation in searches of arXiv and OpenAlex.

AIFS Single v2 adds variables and recovers skill that IFS 50r1 took away. ECMWF's v2 announcement adds wave variables, snow cover, and a 10 hPa level. It reports significant wave height errors about 10% below IFS 50r1, and a 3-day gain in skill at 50 and 100 hPa over AIFS Single v1, with a small negative impact remaining for 2 m temperature in the Arctic. We found no headline v2-versus-v1.1 numbers for 2 m temperature, wind, or radiation. ECMWF's blog on adapting AIFS to IFS cycle 50r1 reports that feeding the 50r1 analyses to AIFS Single v1.1 and AIFS ENS v1 adversely affected them, with degradations of 30% reported for Arctic 2 m temperature, and that fine-tuning on 50r1 data improved the situation. AIFS skill therefore depends on the IFS analysis that AIFS is fed. With both models started from 50r1 analyses, part of v2's lead over v1.1 is recovery from the degradation that 50r1 caused.

Score each AIFS version separately

Each AIFS version should be scored separately, by initialisation time. Each version is a different trained model, so a score pooled across versions describes no single model. The version that produced a forecast is set by the forecast's initialisation time, so a version era is defined by initialisation time, never by valid time. The eras and their dates are:

Era	Initialisation times	Notes
AIFS Single v1.0, first period	from the 2025-02-25 06 UTC run up to, but not including, the 2025-07-31 06 UTC run	First operational AIFS Single
AIFS Single v1.0, after the revert	2025-08-01 (ECMWF states no hour) up to, but not including, the 2025-08-27 06 UTC run	ECMWF reverted to v1.0 after the first v1.1 attempt
AIFS Single v1.1	from the 2025-08-27 06 UTC run up to, but not including, the 2026-05-12 06 UTC run	Runs from 2025-07-31 06 UTC until the revert on 2025-08-01 are excluded (see below)
AIFS Single v2	from the 2026-05-12 06 UTC run	Introduced together with IFS 50r1
AIFS ENS v1	from the 2025-07-01 06 UTC run up to, but not including, the 2026-05-12 06 UTC run	First operational AIFS ENS
AIFS ENS v2	from the 2026-05-12 06 UTC run	Introduced together with IFS 50r1

The v2 boundary coincides with IFS 50r1, so an AIFS change cannot be separated from an IFS change after 2026-05-12. ECMWF introduced AIFS v2 and IFS cycle 50r1 on the same day, and AIFS ingests IFS analyses. A change in the AIFS-versus-IFS score across that date has at least two causes that a score cannot tell apart.

The failed first attempt at AIFS Single v1.1 is excluded, and AIFS Single v1.0 runs on both sides of it. ECMWF implemented v1.1 with the 2025-07-31 06 UTC run and reverted to v1.0 on 2025-08-01 after a stepRange error on six accumulated variables, including ssrd. Forecasts initialised in between belong to neither era, so we drop them. ECMWF re-implemented v1.1 with the 2025-08-27 06 UTC run and reports skill and bias "equivalent (but not identical)" to v1.0 after cutting the soil-moisture loss weight by a factor of 100 to remove spurious point-rainfall artefacts. Even so, the two eras are separate trained models and are scored separately. The dated list of every version is in the NWP model upgrades table, and the journal paper on AIFS Single 1.1.0 (Moldovan et al. 2026) gives the same 2025-02-25 and 2025-08-27 dates.

Reanalyses, hindcasts, and satellite retrievals

Product	Roadmap status	Scored in	Producer and method	Grid	Covers GB	Period	Solar variables	Wind heights	Time step	Latency	Access	Licence	Source
CERRA	🔬 Research (deprioritised)	Past sunshine, on its own row set; wind not yet scored	Copernicus regional reanalysis, HARMONIE-ALADIN with 3D-Var, forced by ERA5	5.5 km, 1069 × 1069 points	Yes (European domain)	1984-09-01 to 2026-06-30 on the CDS on 2026-09-23	Global and direct native, forecast fields only, accumulated	10 m; 15, 30, 50, 75, 100, 150 to 500 m	3-hourly analyses; hourly values from forecast leads 1–3 h (the download holds lead 3 only, as 3-hour accumulations)	About 12 weeks on 2026-09-23	Copernicus Climate Data Store (CDS); no area cropping, so every field is the whole domain	CC BY 4.0	CDS, user guide, Ridal et al. (2024)
COSMO-R6G2	Not on the roadmap	Not scored	DWD, COSMO regional reanalysis forced by ERA5	0.055° (~6 km), EURO-CORDEX EUR-11 domain	Yes	2002-01 to 2025-12	Global and direct native, hourly means	10 m; 100, 150, 200 m	Hourly means stamped at :30	Yearly batches: the 2025 files were posted on 2026-05-11	HTTPS, monthly NetCDF	CC BY 4.0	DWD, paper
ICON-DREAM-EU	Not on the roadmap	Both studies: sunshine, wind	DWD, ICON reanalysis, EU nest	6.5 km	Expected (ICON's EU nest); grid file not checked	2010-01 to 2026-08	Direct and diffuse native, hourly means	10 m; otherwise model levels only, so a fixed height needs vertical interpolation	Hourly	The August 2026 file was posted on 2026-09-07 (DWD states 2–3 months)	HTTPS, monthly GRIB2	CC BY 4.0	DWD
NORA3	Not on the roadmap	Not scored	MET Norway, HARMONIE-AROME hindcast downscaling ERA5	3 km	Yes, checked on the grid	1958 to 2026-08-31	Global only	10 m; 20, 50, 100, 250, 500, 750 m, as grid-relative components needing rotation	Hourly, from leads 3–9 h of each 6-hourly run	About 3 weeks	THREDDS / OPeNDAP	CC BY 4.0	THREDDS, Solbrekke et al. (2021)
MERRA-2	Not on the roadmap	Not scored	NASA's Global Modeling and Assimilation Office, Goddard Earth Observing System (GEOS) global reanalysis	0.5° × 0.625°	Yes (global)	1980 to present	Global only (`SWGDN`)	2, 10, 50 m	Hourly means stamped at :30	About 3 weeks after month end	GES DISC, Earthdata login	NASA open data	radiation, wind
ERA5-Land	Not on the roadmap	Not scored	ECMWF, land component of ERA5 re-run with ERA5 forcing	0.1°	Yes	1950 to 2026-09-17	Global only	10 m only	Hourly	About 6 days	CDS	CC BY 4.0	CDS
CERRA-Land	Not on the roadmap	Not scored	Copernicus land-surface reanalysis on CERRA's grid	5.5 km	Yes (European domain)	1984-09 to 2025-12-31	Global only	None	Leads 1–3 h	Not checked	CDS	CC BY 4.0	CDS
NOAA GFS and GEFS analyses (Dynamical.org)	Not on the roadmap	Not scored	The first hours of each GFS or GEFS run, concatenated	0.25°	Yes (global)	GFS 2021-05-01; GEFS 2000-01-01 (reforecast to 2019, operational after)	Global only	10, 80, 100 m	GFS hourly; GEFS 3-hourly	Not established	Icechunk (Zarr)	Public domain from NOAA; CC BY 4.0 on Dynamical.org	GFS, GEFS
HelioClim-3	Not on the roadmap	Not scored	Transvalor and Mines Paris – PSL, Heliosat-2 retrieval from Meteosat	About 5 km in Europe	Yes	2004-02 to real time	Global; direct and diffuse by decomposition	None	15 min	"A few minutes after image reception"	SoDa web service	Paid after 2006	SoDa
Solcast	Not on the roadmap	Not scored	Solcast, multi-satellite retrieval	90 m (downscaled)	Yes	2007-01 to 7 days ago	Global, direct normal, diffuse, tilted	Included; heights not listed	5–60 min	7 days in the historical API; Solcast's live API is separate	Commercial API	Paid	Solcast

"Latency" in this table is how far the archive lags real time.

Weather-station observations

The studies also read one source of station observations, the Met Office's MIDAS Open, which has no row in the tables above because it is a set of points rather than a gridded product. MIDAS is the Met Office Integrated Data Archive System. MIDAS Open is its open release, published by the Centre for Environmental Data Analysis (CEDA) under the Open Government Licence v3.0. The release read here is dataset-version-202607, quality-control version 1. The release covers 2017-01-01 to 2025-12-31. CEDA releases MIDAS Open once a year, so the files hold only past weather. The catalogue lists no station source, and the project ingests none. See the catalogue's summary of what the station arms found.

The studies downloaded 10 of the 86 radiation stations whose records overlap their years, and 38 hourly-weather stations. The lists were typed by hand, and one distance rule reproduces them: a station is included if its record runs into 2025 or later and it lies within 100 km of at least one of the nine anonymised study sites. That rule gives the 10 radiation stations and 37 of the 38 hourly-weather stations. The 38th station's record ends in 2024, and the download holds it although it fails the record test. The other 76 radiation stations are all 107 km or more from every solar farm. See how the stations were chosen.

The variables are hourly global irradiation, 10 m wind, and air temperature.

Global horizontal irradiation is an hourly total in kJ m⁻² over the hour ending at the timestamp, in UTC. Dividing by 3.6 gives the hour's mean irradiance in W m⁻². The diffuse and direct columns are empty at all 10 stations.
10 m wind speed and direction are a 10-minute mean over the window from 20 to 10 minutes before the timestamp ("HH-20 to HH-10" in the Met Office's notation), with speeds in whole knots and directions in steps of 10 degrees. Only 18 of the 38 stations report an hourly wind speed. The wind study did not find the averaging window documented for the 4 stations that report the automatic-weather-station hourly message type.
Air temperature is a spot reading at the timestamp, reported at all 38 stations. Some stations report only once a day.

Measured against gridded products, a station's irradiance scores between CAMS and ERA5 on sunshine, and a station's 10 m wind scores worse than ERA5's on wind. For six solar farms in Lincolnshire, all sharing one radiation station 17 to 31 km away, the station's irradiance and temperature score worse than CAMS and better than ERA5, and adding the station's irradiance to CAMS lowers CAMS's error. The figures and intervals are in Met Office weather-station observations as a stand-in for a gridded product. For three wind farms, each farm's nearest anemometer, 6 to 18 km away, gives a larger error than ERA5's 10 m wind, and lowers UKV's error when added to it. The figures are in One nearby 10 m weather station trails ERA5's 10 m wind on its own, and lowers UKV's error when added to it.

Every station result shares four limits.

The data end on 2025-12-31. The station arms therefore score fewer hours than the main arms and end before the Met Office's UKV upgrade of January 2026. See what the station files hold, and where they end.
The quality-control flags are kept as delivered, and no row is dropped on a flag. Flag 106 marks whole stations or runs of several months rather than isolated bad hours. The wind study did not check whether excluding a flagged station moves its contrasts, and the solar study checked only the 107 site-hours whose flag differs from the usual value.
Each planned contrast reads one nearest station per farm. The six solar farms share one radiation station, so the results do not say how a different station would compare.
MIDAS Open's other wind datasets were not searched for the 12 hourly-weather stations with no wind in the downloaded file, so a nearer anemometer may exist. See the weather-station arms.

The studies did not research whether a crowd-sourced observation dataset, such as a network of amateur weather stations, would add anything to a station arm.

Candidates for the past-sunshine study

Of the products surveyed, ECMWF IFS HRES 9 km, CERRA, and COSMO-R6G2 rank highest for the past-sunshine study, because each adds a model family, a grid spacing, or a kind of product the study has not scored over the whole window from December 2022. A native direct beam counts for less: the study found that a product's own direct beam improves its mean absolute error by 0.03 to 0.10 percentage points of capacity. A place in this ranking is not a decision to score the product, and no extension of the study has been agreed. The ranking, with each product's roadmap status and one reason:

ECMWF IFS HRES 9 km on Open-Meteo (not on the roadmap) — a 9 km forecast from the IFS model behind the 31 km ERA5, from 2017, through the same Open-Meteo interface as the four weather models on the main row set, with a native direct beam. The study scores it only from November 2024, on the shorter row set of the four extra Open-Meteo models, so the candidate is to score it over the whole window from December 2022. Before October 2025 the series is Open-Meteo's own assembly of IFS runs "employing the most up-to-date version of IFS" (Historical Weather API). Open-Meteo does not document the series' served lead. A study using the series has to report both the assembly and the undocumented lead.
CERRA (🔬 research, deprioritised) — a regional reanalysis on a 5.5 km grid with a native direct field, hourly through forecast leads 1 to 3 (the download holds lead 3 only, as 3-hour accumulations, and the past-sunshine study scores CERRA from them). Its latency is why the catalogue deprioritises CERRA for the live service, and is no obstacle for a study of past weather. The download volume is the obstacle: every field is the whole European domain.
COSMO-R6G2 (not on the roadmap) — a second regional-reanalysis family, after CERRA, on a ~6 km grid, with hourly global and direct irradiance as plain monthly NetCDF. The archive stops at December 2025, 9 months short of the study's September 2026 end, so COSMO-R6G2 can only be scored on a shorter window. DWD publishes the archive in yearly batches. The hourly means are stamped at :30 and need shifting 30 minutes to be period-ending.
GFS (0.11° surface fields) (not on the roadmap) — a further global centre, from 2021-03-23, whose diffuse field is visible-band rather than broadband.
MERRA-2 (not on the roadmap) — an optional arm from the GEOS model family. The ~50 km grid is coarser than every other arm, so a score cannot separate model family from grid spacing, and MERRA-2 has no direct field. The hourly means are stamped at :30 and need shifting 30 minutes to be period-ending.
HelioClim-3 (not on the roadmap) — a second satellite retrieval beside CAMS, but paid after 2006, and its direct and diffuse are decompositions of global irradiance.

ICON-DREAM-EU, an ICON reanalysis on a 6.5 km grid with native direct and diffuse, is scored in the past-sunshine study: ICON-DREAM-EU beats ERA5 but not the ICON weather models.

The sunshine study scores the DMI and KNMI HARMONIE-AROME forecasts on global irradiance only, on its shorter row set from November 2024. It scores neither direct beam. KNMI's direct beam is Open-Meteo's separation of the global irradiance. The study finds that the field Open-Meteo serves as DMI's direct beam is exactly zero in half of the daytime rows and in some rows exceeds the global irradiance, which supports the reading that the field is not DMI's direct beam.

Candidates for the past-wind study

Of the products surveyed, CERRA and NORA3 rank highest for the past-wind study, because both serve wind close to hub height on grids of 3 to 5.5 km. Neither covers the whole study window: CERRA ends on 2026-06-30 and NORA3 on 2026-08-31, against the study's September 2026 end. The wind farms' hub heights are not known; the study scores 80 m and 100 m wind. A place in this ranking is not a decision to score the product, and no extension of the study has been agreed. The ranking, with each product's roadmap status and one reason:

CERRA (🔬 research, deprioritised) — wind at 75 m and 100 m, bracketing 80 m, on a 5.5 km grid, hourly, to June 2026. The download volume is the same obstacle as for sunshine.
NORA3 (not on the roadmap) — HARMONIE-AROME on a 3 km grid with wind at 50 m and 100 m, hourly, to August 2026, over OPeNDAP, so a point extraction is cheap. The components are grid-relative and need rotating.
COSMO-R6G2 (not on the roadmap) — 100 m wind, hourly, on a ~6 km grid, to December 2025.
ECMWF IFS HRES 9 km (not on the roadmap) — 10, 80, and 100 m wind from 2017, through the existing Open-Meteo interface.
DMI HARMONIE-AROME DINI (not on the roadmap) — 100 m wind on a 2 km grid from 2024-06-28, from a different consortium than UKV and ICON-D2. The 80 m wind Open-Meteo serves is interpolated.
GFS (not on the roadmap) — 80 m and 100 m wind, but only from the 0.25° product.

MERRA-2 earns no wind arm: its highest wind is at 50 m, below both heights the study scores, on a ~50 km grid over flat Lincolnshire, so a score would mainly measure grid spacing.

ICON-DREAM-EU is scored in the past-wind study: ICON-DREAM-EU does not beat ERA5, and trails ICON-EU. The study reads it at DWD's model level 72, about 96 m, directly, rather than interpolating to a fixed height.

Candidates for comparing forecast models at a fixed lead (issue #810)

Several whole-run archives reach back further than Open-Meteo's Previous Runs archive, but Previous Runs is the only free route surveyed that puts ICON-D2, AROME, the HARMONIE-AROME models, and UKV as Open-Meteo serves it at a fixed lead, and for those models at day 1 only. CEDA's UKV archive keeps whole runs from 2016, but access is by application, and CEDA's UKV is statistically different from the live UKV. Issue #810 proposes ECMWF ENS via Dynamical.org, the product in production, as the incumbent each other model is compared against. A place in this list is not a decision to score the product. Ordered by how far back whole runs reach:

Dynamical.org GEFS 35-day (not on the roadmap), from 2020-10-01 — an ensemble comparator to ECMWF ENS with 3.5 more years of history, 00 UTC runs only.
Dynamical.org GFS (not on the roadmap), from 2021-05-01 — 4 runs a day to 384 h with global irradiance and wind at 10, 80, and 100 m.
Google WeatherNext 2 (not on the roadmap), from 2022-01-01 — an ensemble with wind and no radiation, marked experimental by Google.
Open-Meteo Single Runs of ECMWF IFS HRES 9 km (not on the roadmap), from 2024-03-14 — the one whole-run archive in this list with a native direct beam. Three caveats apply: Open-Meteo calls the early runs hindcasts, the runs sampled from early 2024 are 00 and 12 UTC only, and the model changes to IFS Cycle 50r1 on 2026-05-12, an upgrade the data sources page describes.
Dynamical.org AIFS Single (not on the roadmap), with radiation and 100 m wind from the 2025-02-24 06 UTC run.
Open-Meteo Previous Runs, from 2024-01-19 at the earliest — capped at the offsets in the table above.

The matched-lead study scored most of these candidates and found ENS ahead of UKV, ICON-EU, and GEFS for solar. The study reads Open-Meteo Previous Runs (UKV, ICON-D2, ICON-EU, ICON global, IFS 0.25°, GFS, ARPEGE, AROME, and two HARMONIE-AROME products) at whole-day offsets, and Dynamical.org's ECMWF ENS and GEFS as whole 00 UTC runs. Previous Runs is not a fixed-lead archive, because a value's lead depends on each product's run cycle, so the study brackets each product between ENS's day-0 and day-1 errors. For solar, UKV (+1.242 points [+0.824, +1.689]) and ICON-EU (+0.817 [+0.584, +1.069]) lose to ENS, and GEFS at an identical lead is +1.272 points [+0.937, +1.601] worse. For wind, UKV loses (+0.770 [+0.460, +1.057]), ICON-EU is unresolved (+0.169 [-0.064, +0.384]), and GEFS is +0.858 points [+0.598, +1.099] worse. Dynamical.org's GEFS is read from 2024-11-30 in the study, although the archive reaches back to 2020-10-01. Products the study does not fit are the ERA5 and CAMS reference rows, IFS HRES 9 km, Dynamical.org's own GFS, and Open-Meteo Single Runs.

Dynamical.org's ICON-EU whole runs carry native direct and diffuse irradiance, but start only on 2026-02-10, too short a history to list.

AI weather models beyond AIFS

Of the AI weather models with an open or free-on-request data feed that we checked on 2026-09-25, other than AIFS, only WeatherNext 3 carries both solar radiation and 100 m wind, and its archive starts on 2026-01-01. This section lists the other AI models that could feed a forecast of substation load, and the commercial providers that sell AI forecasts or archives of them. The search was not exhaustive, and an absence claim below means "not found in what we searched". Vendor accuracy claims are the vendor's own and are not repeated here. We have no score comparison between any of these models, so the section says nothing about which is more accurate. The forecast-model table already covers AIFS Single, AIFS-ENS, GFS, GEFS, and WeatherNext 3, and this section does not repeat those rows. "The weather-product studies" here means the past-sunshine study and the past-wind study.

Open AI forecasts from DWD and NOAA carry no radiation, and little 100 m wind

Model	Status	Radiation	100 m wind	Archive of past runs	Licence	Source
DWD AICON (Deutscher Wetterdienst)	Operational since 2026-03-02; 13 km, 3-hourly to 48 h then 6-hourly to 180 h	None in the open-data listing	No height-level field; 10 m wind, plus `U` and `V` on native model levels (per the paper)	None: 4 runs kept, about 1 day	CC BY 4.0 (DWD open-data terms, linked from the server's README)	Paper, open data
NOAA AIGFS	Operational; 4 runs a day, 6-hourly steps to 384 h	None in the surface file	From the 2026-07-30 00 UTC run; absent from earlier folders	Bucket folders from 2026-04-16; re-runs from 2025-01-01 in `hindcast_aigfsdev2.5/`	CC0 (AWS registry)	NOAA registry entry, bucket
NOAA GraphCast-GFS (experimental predecessor of AIGFS)	Bucket folders from 2024-02-05 to 2026-05-05	None (surface file, 2026-05-04 00 UTC run)	None (10 m only)	Bucket folders, and GribStream from 2024-04-25	CC0 (AWS registry)	NOAA registry entry
AIWP reforecasts (NOAA and CIRA, the Cooperative Institute for Research in the Atmosphere)	Research reforecast archive of FourCastNet v2-small, Pangu-Weather, and GraphCast, started from GFS and IFS analyses; 00 and 12 UTC runs to 240 h, 6-hourly	None listed	FourCastNet v2-small only	From 10/2020 (FourCastNet, Pangu-Weather) or 01/2022 (GraphCast)	"Open Data. There are no restrictions on the use of this data"	AWS registry

AICON cannot serve either of the weather-product studies, and its open-data listing keeps only the latest four runs. The listing at opendata.dwd.de/weather/nwp/v1/m/aicon/p/ holds 12 variables. Among them are T_2M, U_10M, and V_10M, and none is a radiation field. The only other wind variables are U and V, and the paper says output is stored on native model levels. The T_2M directory held the runs of 2026-09-24 12 UTC to 2026-09-25 06 UTC when we listed it, so a backtest cannot be built from DWD's server. The paper describes an operational model trained on DWD's ICON-DREAM reanalysis, with a re-forecast for March to September 2025 used only for evaluation.

AIGFS acquired 100 m wind in the surface file during 2026 and has no radiation field. The 2026-04-16 folder holds UGRD and VGRD at 10 m, TMP at 2 m, PRMSL, and APCP. Files from the 2026-09-24 00, 06, 12, and 18 UTC runs add 100 m UGRD and VGRD and 2 m DPT. We checked the 00 UTC run's 24-hour surface index in each folder, and the first folder with 100 m wind is aigfs.20260730, so the archive holds about 8 weeks of 100 m wind. NOAA's registry entry says the AIGFS, the AI Global Ensemble Forecast System (AIGEFS), and the Hybrid Global Ensemble Forecast System (HGEFS) became operational on 2025-12-17, and GribStream's blog gives the same date, starting with the 12 UTC cycle. The bucket's earliest aigfs.* folder is 2026-04-16, and that folder holds only the 06 UTC run. The next folder is 2026-04-21. AIGEFS is under EAGLE_ensemble/aigefs.*, from 2025-06-01 to 2026-09-25, and its surface file has 10 m wind and no radiation or 100 m wind. The bucket also holds hindcast_aigfsdev2.5/, re-runs from 2025-01-01 to 2026-02-28 that carry 10 m wind and no radiation or 100 m wind.

The AIWP archive has no radiation for any of its three models and 100 m wind for one. FourCastNet v2-small is the only model in the listing with 100 m wind components. The archive is a research reforecast and not a set of as-issued operational forecasts, and its own page says data "may be missing and is not guaranteed to be available at any given time".

Three more AI models publish weights or research results, and none has a public forecast feed we found

Model	What is published	Source
ECMWF AIFS-DOP (Direct Observation Prediction)	Research paper: a model trained on 40 years of gridded observations alone, on an O96 grid of about 100 km with 6-hour slices, scored against radiosondes aloft and SYNOP surface observations, including 2 m temperature and 10 m wind. The paper does not present it as operational, and we found no public forecast data	Paper
ECMWF machine-learned reanalysis prototype (made with AIFS-DOP)	ECMWF's news item says a prototype reanalysis for 1981 to 2022, on a grid of about 112 km every 6 hours, was produced "in the course of a single working day". It is a proof of concept	ECMWF
Met Office and Alan Turing Institute FastNet	Model weights for inference only, under the Open Government Licence v3.0, on an O96 grid of about 104 km; described as the first experimental release	Hugging Face
NVIDIA Earth-2 (Atlas, StormScope, HealDA)	Announced on 2026-01-26 as open models, to be licensed for commercial and noncommercial use and released on GitHub and Hugging Face. We found no hosted forecast archive	SiliconANGLE

Running a published model ourselves would produce forecasts that no weather service issued. A model with public weights gives us a forecast only for the days we run it, and re-running past days starts from an analysis chosen by us. Whether a re-run counts as a forecast for backtesting is the question the next section asks of commercial archives. Searches for public AI feeds covering Great Britain from Météo-France, KNMI (the Royal Netherlands Meteorological Institute), MeteoSwiss, DMI (the Danish Meteorological Institute), Environment and Climate Change Canada, and the China Meteorological Administration found none. Dynamical.org's catalogue lists AIFS Single and AIFS-ENS and no other AI model.

For backtesting, the deciding property is an archive of as-issued forecasts with radiation and 100 m wind

The question to ask any provider is whether "historical forecast" means the runs as they were issued, or a re-analysis or a re-run. A series stitched from successive one-day forecasts, or re-run from a later analysis, is not the forecast a substation-load model would have received on the day. The rows below say what each provider's own page states. Only Prescient states both an as-issued archive and radiation with 100 m wind, and its coverage of Great Britain is unclear.

Provider	Product	Archive of past forecasts, as the page states it	Notes	Source
Jua	EPT-2 model family (0.081° global, 4 runs a day, 20 days), plus third-party models including AIFS	Docs call all archives "hindcast datasets" and list EPT 2 data from 2023-01-01. The page does not say whether the runs are as issued	Global and direct shortwave flux; wind at 10 to 200 m; price unpublished. EPT 2.1 Helios is a solar-only 30-minute nowcast served over Europe from 2024-01-01, and EPT 2.1 Europa is a 16-member, 48 h hourly ensemble over Europe from 2025-08-14. Jua also lists hindcasts of third-party models, including AIFS and Aurora	Jua docs
Brightband	Archive of forecasts from GraphCast, Pangu-Weather, and AIFS-Single, all initialised from IFS HRES	2021 to 2024, 280 TB, in Zarr with Icechunk on Google Cloud, with a subset on Arraylake that reproduces the ExtremeWeatherBench benchmark. Models published after 2021 can only have been run retrospectively, so we read these as re-runs, not as-issued forecasts	Brightband says it runs Aurora, Earth-2 Medium Range/FourCastNet-v3, and AIFS-ENS and will add them to the archive, but the page does not name AIFS-ENS as part of it today, and states no licence, grid, or variable list; none of these is verified. We could not list the Google Cloud bucket without credentials, so public access is unverified. The archive stops before most of our study window. Its marketplace listing is not among the 24 datasets visible on the Earthmover marketplace on 2026-09-25, so the route to the full archive is unverified	Brightband, marketplace
Meteomatics	REST API with AIFS (variant not stated)	"Historical forecasts from 2017", and ERA5 from 1940	The page says "Meteomatics purchases and saves historical weather forecasts", which supports an as-issued reading that the page does not confirm. It does not say which variables AIFS carries	Meteomatics, AIFS
Prescient Weather (World Climate Service)	Point-in-time archive of AIFS-ENS, AIFS-Single, ECMWF, GEFS, and GFS, aggregated to country and region level	"9+ years of as-issued history"	Capacity-weighted 100 m wind and solar radiation indices. The archive's start date differs by model, since AIFS did not exist for 9 years. The archive page lists "European Union: Country Level" and the point-in-time page names "UK NBP" (the National Balancing Point gas market), and the UK is outside the European Union, so whether Great Britain is a region is not established	Prescient, World Climate Service
meteoblue history+	NEMS global model, ERA5, and local models	Each forecast is "archived by meteoblue at least once a day", and the history is a continuous hourly series built from one-day forecasts	A stitched series, not whole past runs	meteoblue
GribStream	GraphCast-GFS through the 2026-05-05 run, and other models	From 2024-04-25	Radiation not listed; the model is the experimental predecessor of AIGFS	GribStream
Solcast	Irradiance, wind, and weather data	Historical time series from 2007 to 7 days ago; forecasts from now to 14 days	No archive of as-issued forecast runs found. Free historical data for researchers at public research institutions, for non-commercial use, through 50 free requests	Solcast
Zeus AI (Helios)	Satellite-based global horizontal, direct normal, and diffuse horizontal irradiance, 6 km, hourly	The free marketplace listing "Helios Solar Irradiance - Historical (Americas)" is a static dataset from 2020-01-01 to 2023-12-31. Real-time actuals and 6-hour forecasts are available by contacting Zeus AI	The listing covers 80°S to 80°N and 5°W to 150°W, so it does not include Great Britain. It has no forecast archive	Zeus AI, listing
Earthmover Datasets	"ECMWF IFS HRES Archive": ECMWF's HRES surface forecasts at the native 0.1° (about 9 km), curated from ECMWF's MARS archive as an Icechunk cube	Forecasts initialised from 2023-01-01 00 UTC to the present, at 00, 06, 12, and 18 UTC, refreshed daily. Whether the runs are as issued is not stated, though the listing says the archive "faithfully reflects what ECMWF produced"	12 surface variables, with 10 m and 100 m wind, 2 m temperature and dewpoint, total and convective precipitation, mean sea-level pressure, snow depth, downward surface solar radiation, and surface runoff. No direct-beam field is listed. Lead times run to 360 h for 00 and 12 UTC runs after the Cy49r1 upgrade (operational 2024-11-12 12 UTC), to 240 h for earlier 00 and 12 UTC runs, and to 90 h for 06 and 18 UTC runs. Steps are hourly to 90 h, 3-hourly from 93 to 144 h, and 6-hourly from 150 to 360 h. Paid: the price is shown only after login, and the licence tab reads "Earthmover commercial data license, available on request". The listing does not say what it holds for runs after ECMWF's 2026-05-12 change	listing
Beyond Weather	A fine-tuned AIFS optimised for 100 m wind	Not stated	Terms not verified	Earthmover
Silurian (Generative Forecasting Transformer, 1.5 billion parameters)	A research paper, not a data feed	Not applicable	Fine-tuned on Hydro-Québec's observations and tested on 65 wind-farm masts from January 2024 to January 2025 and on 534 temperature stations to March 2025, all in Québec. It is not a test in Great Britain	Bodnar et al. (2025)
Microsoft Aurora	Model on Azure AI Foundry	Hindcasts listed by Jua and run by Brightband; none found from Microsoft	Hosted endpoint for running the model	Aurora documentation
Huawei Pangu-Weather	Weights	Not applicable	Weights under CC BY-NC-SA 4.0, which forbids commercial use	GitHub

Three of these providers state an archive that could serve a backtest, and each has a gap. Jua's archive starts in 2023 but its page does not say whether the runs are as issued. Brightband's archive stops in 2024 and, we read, consists of re-runs. Prescient's archive is as issued but aggregated to regions, with the coverage of Great Britain unclear. Meteomatics says it saves historical forecasts from 2017, without saying whether they are as issued.

Earthmover's "ECMWF IFS HRES Archive" differs from Open-Meteo's IFS HRES 9 km archives in start date, cycles, grid, and price. Its runs start on 2023-01-01, against 2024-03-14 for Open-Meteo Single Runs, 2025-10-01 for Open-Meteo Previous Runs at 9 km, and 2024-03-06 for Open-Meteo Previous Runs of IFS 0.25°. It holds four cycles a day throughout, where Open-Meteo Single Runs began with 00 and 12 UTC and added 06 UTC on 2024-08-10. It serves the native 0.1° grid, and its 12 variables include downward surface solar radiation but no direct-beam field, where Open-Meteo serves ECMWF's direct irradiance natively. The price is unknown, where Open-Meteo is free for non-commercial use. This survey does not assess buying it.

Open-Meteo's own documentation and the survey give different start dates for AIFS archives, and the survey's measured date governs. For AIFS, Open-Meteo's Previous Runs page says "Most models are archived from January 2024". The forecast-model table gives 2025-02-17 for AIFS Single, the first day with global irradiance at Lincoln, and that measured date governs.

Earthmover's data marketplace is a delivery route for several of these archives, and its listing page can be read without a login. The listing page showed 24 distinct datasets on 2026-09-25, with no pagination and no documented API or machine-readable index. Fetching a listing URL shows its README, dataset, and licence tabs, but subscribing needs a login. Any Arraylake account can subscribe instantly to a free listing. Paid listings are set up with the provider, and Earthmover "does not sell the data directly" (Earthmover documentation). The launch providers were Brightband, Dynamical.org, Zeus AI, Beyond Weather, Planette, the Climate Uncertainty Lab, and the Subseasonal Consortium (Earthmover).

On 2026-09-25 the marketplace listed datasets from seven providers. Dynamical.org has 19 (the GFS, GEFS, IFS ENS, AIFS Single, AIFS ENS, ICON-EU, HRRR, MRMS, and IMERG products and analyses that this survey already covers). Brightband has 2 free listings, "IFS 15-day forecast (open)" and "IFS ensemble initial conditions (open)". Spire has 1 free listing, an S2S hindcast of its SAIFS model at 0.5°. The Climate Uncertainty Lab has 2 free bias-corrected CMIP6 temperature datasets. CTrees has a biomass dataset, Zeus AI has 1 free irradiance dataset (Americas only), and Earthmover Datasets has 1 paid listing, the HRES archive above. Some entries appear twice as featured datasets. We found no listing from Jua, Tomorrow.io, Salient, WindBorne, Excarta, Beyond Weather, or Silurian.

We found nothing on archives, variables, or AI models from eight more providers: Xweather and Vaisala, DTN, Vortex, Meteocontrol, Amperon, Nnergix, Enfor, and Meteologica. One Meteologica page returned an access error.

Products ruled out

The products ruled out each fail on period, on missing variables, on having no data at Lincoln, or on adding nothing a product already on this page carries:

UERRA — ends in 2019, and CERRA supersedes it (CDS).
COSMO-REA6 — ends 2019-08-31, and COSMO-R6G2 supersedes it (WDC Climate).
ERA5-Land — its global irradiance and 10 m wind come from ERA5's forcing, per the CDS description, and it has no direct field and no hub-height wind.
CERRA-Land — global irradiance only, no wind, and ends 2025-12-31.
The National Solar Radiation Database (NSRDB), Meteosat Prime Meridian — covers Europe, but ends in 2022 (NSRDB).
Met Office IMPROVER Blended Probabilistic Forecast on AWS — a calibrated blend of percentiles and probabilities, with total and low cloud but no irradiance, and wind at 10 m only, in the UK gridded buckets. The BPF holds no raw MOGREPS-UK members, and the buckets keep 30 days (details).
Open-Meteo Ensemble API as a history source — values at Lincoln reach only 3 to 4 days back, and less than a day for the ECMWF 9 km Europe ensemble, although the API accepts dates 92 days back (Ensemble API).
ECMWF IFS ENS 46-day — 1.5° grid and no radiation (Dynamical.org).
Google WeatherNext Graph — deprecated, with 10 m wind only and no radiation (Earth Engine).
ECMWF IFS 0.4° (ecmwf_ifs04) on Open-Meteo — no radiation, 10 m wind only, and the series stops 2025-03-07.
Météo-France AROME France HD on Open-Meteo — no radiation at all.
Open-Meteo "UK Met Office seamless" — at Lincoln on 2023-06-10 and 2025-06-10 every global irradiance and 10 m wind value equals UKV's.
MET Nordic, JMA MSM, ItaliaMeteo ICON-2I, MeteoSwiss ICON-CH1, and KNMI HARMONIE-AROME Netherlands on Open-Meteo — no data at Lincoln.
ACCESS-G on Open-Meteo — no radiation or 100 m wind at Lincoln on 2026-09-01.
GEM Global and CMA GRAPES on Open-Meteo — data at Lincoln, but neither grid spacing nor method sets them apart from the global models ranked above.
MOGREPS-UK and MOGREPS-G history — AWS holds a 30-day rolling window (MOGREPS-G), and CEDA's ukmo-nwp archive has no MOGREPS directory. The deprecated registry.opendata.aws/mogreps/ listing is still readable but covers only 2013 to 2016, ending nearly three years before NGED's power data begins. On 2026-09-26, Open-Meteo's ensemble API held the MOGREPS-UK members for about 3.5 days and its ensemble-mean API held the mean and spread from 2026-06-25. What we learnt about MOGREPS-UK and what we learnt about MOGREPS-Global have the detail.