Weather inputs by product

Each weather product reaches the forecast at a different time, and an experiment must name which product feeds each column at forecast time and which feeds it at training time. The two are usually different products. This page is a digest: it holds the table of what exists when, and the traps that cost experiments a re-run. The long accounts are the weather products survey, the data-sources page, the NWP variable conventions, and the ECMWF ENS known issues.

Availability at forecast time

Product	What it is	Delay before use	Exists at forecast time?	Training rule
ECMWF ENS (51 members)	Ensemble forecast, stored by `delta_store.nwp`	A run is usable 9 hours after its `init_time` (`NWP_PUBLICATION_DELAY_HOURS`)	Yes: the production input	Train on ENS history, which starts 2024-04-01 on disk
ERA5	ECMWF reanalysis, an estimate of past weather	About 5 days through the Climate Data Store	No: a ceiling at lag times	Version-1 plan: the training product for every variable except irradiance
CAMS solar radiation	Satellite-derived irradiance, point time-series API	1 day: a request that names today is rejected	No: nothing about the day being forecast	The planned training product for irradiance
UKV (Met Office)	2 km deterministic UK weather model	Three variants exist; see below	Only the live feed	Train on the product that will serve; never train on the Centre for Environmental Data Analysis (CEDA) archive and serve the live feed
GEFS, ICON-EU, IFS, AIFS-ENS, WeatherNext 3	Further forecast products	Differs by product	Studies only	Not in production; see the weather theme

A weather product that is "as accurate as ERA5" is a ceiling for demand, not a forecast. ERA5 weather in place of ENS weather lowered the Primary demand quantile score by 0.10 points (exp 22; Primary demand, one validation year; provisional).
Name the weather source used at lag times in every write-up, and say whether it exists at forecast time (data-processing page; exp 33, 37; ERA5 as the lag-time source; established).

Training product: the rule

Train each XGBoost model on the history of the weather product that will serve it. The arms are ENS-trained, ENS plus the other product as columns, and the other product alone. An ERA5-trained XGBoost model is only a reference (weather theme; exp 22; XGBoost models on Primary demand; provisional).
Pair a CAMS-trained XGBoost model with a stage 0 quantile mapping of ENS shortwave. The pair scored 0.102 points better than the ERA5-trained arm for Primary demand, against 0.035 without the mapping (exp 14; Primary demand; provisional, and it supersedes the advice of exp 9 to keep ERA5).
Calibrate ENS temperature and irradiance before they drive a model trained on past weather. Shifts of up to about 2 K and 20% were needed at five Primaries (exp 111; 5 Primaries; provisional).
"Better weather adds little" was measured only for an XGBoost model trained on ENS. Do not read it as a result for an ERA5-trained XGBoost model (weather theme; exp 22; ENS-trained XGBoost; provisional).
Training on ERA5 lowered the solar and wind farm error up to day 6, but for demand a long history did not carry over (weather theme; exp 1, 3; 9 generator series, 22 demand series; provisional).

ECMWF ENS

The step grid and the publication delay are owned by the architecture pages. See the step grid and the ECMWF ENS known issues (code).
The ENS archive on disk starts at 2024-04-01. Source Cooperative stages about 3 further years as raw GRIB1 (complete dates 2021-03-21 to 2024-03-31), and a pilot decoded them bit-exact. See the training-history page for whether the full fetch has run (disk; docs).
Instantaneous and period-ending variables differ. See the NWP variable conventions (code).
Drop an ensemble member that has a missing weather value, and refuse a run whose control member is incomplete. studies.dp_ens_weather does both (code; exp 111 dropped one member of one run).
Repeated ENS runs repeat each valid time about 15 times. An ENS-trained model with calendar columns can memorise the calendar position, so include a control without them and group inner splits by valid time (exp 1, 3; ENS-trained XGBoost; provisional).

ERA5

ERA5 arrives about 5 days late through the Climate Data Store. A faster public Zarr copy (ARCO-ERA5) carries ERA5T at about a week's lag (data-sources page) (docs).
The half-hourly file's :30 values are interpolated. Each is the mean of the two neighbouring hourly values and contains the next hour's reading (exp 104; ERA5 half-hourly file; provisional).
ERA5T hours from the last few months are preliminary. See the five traps (docs).
ERA5 is one of several estimates of past weather, and it describes past sunshine and wind worse than several current products. CAMS beat ERA5 by 3.7 to 4.1 points of capacity on past sunshine at the solar study's five headline row sets, an exploratory comparison (training-history page; solar study, 5 row sets; provisional).

CAMS

Use the point time-series API, not the gridded product. The gridded product lags years behind (data-sources page) (docs).
The latency is 1 day. A forecast job at 09:00 UTC gets nothing about the day it forecasts, so no intraday correction rests on CAMS (docs).
Values are accumulated irradiation in Wh/m² and each is stamped at the end of its step. Set time_reference to universal_time (docs).
Snapshot each fetch. CAMS recomputes a request against the latest software, so the same timestamp can return a different value later (docs).

UKV

Three UKV variants are different products. The CEDA archive, Open-Meteo's archive, and the live feed differ in lead, irradiance, temperature, and units (data-sources page) (docs).
Open-Meteo's UKV archive is the T+0 analysis. It holds no forecast at any lead, so it compares with a reanalysis and does not measure forecast skill (docs).
The CEDA archive has no 100 m wind, lacks four hourly steps, and is licensed for non-commercial use. UKV was upgraded on 2026-01-21 (exp 54, 57; CEDA archive; provisional).
UKV shortwave is instantaneous, whereas ENS and CAMS are means labelled at the period end. Check the hour-of-day peak of every source (exp 53; UKV, ENS, CAMS; provisional).

Findings on the weather inputs of generator forecasts

A UKV forecast adds 1.1% to 1.8% to the solar-farm score. Experiment 229 found 1.8%, and experiment 248 found 1.1% for the neural head and 1.4% for XGBoost. These gains are upper limits, because they use the research archive and the live UKV feed differs. ICON-D2 adds 1.0% in a blend at 12 to 29 hours only, and experiment 246 advises against adding it (exp 229, 246, 248; 6 farms, 10 months; provisional).
The ECMWF ensemble's wind speeds are about 1.1 m/s too slow, which gives the low median. Multiplying every member's wind speed by a factor fitted on the 14 months before the scored year closed about 78% of the gap to the ERA5 oracle, which is the same physics core fed ERA5 wind as if it were the forecast. A lookup table of power offsets by lead day helped more than mapping the members to ERA5 (exp 261, 285, 307; 3 wind farms; provisional).
The 51 ensemble members give better wind tails than their mean. Pooling the AIFS ensemble (ECMWF's machine-learned ensemble) lowers the wind score at days 3 to 6 by about 5% and does not improve the high-power tail (exp 242, 262; 3 wind farms; provisional).
Extra ECMWF ensemble variables and the cells around a farm add under 0.4% for solar farms. The variables were longwave radiation, humidity, precipitation, pressure, wind, and 500 hPa height; the neighbouring cells improved the score by 0.36% (exp 272, 292; 6 farms; provisional).

Other products

GEFS did not improve the sunshine anomaly, and its raw anomaly was about 15% too large (exp 50; sunshine anomaly; provisional).
ICON-EU and IFS winds scored worse than the best guess when fed alone, as raw winds with no bias correction, to an ERA5-trained wind model: 2.5 and 1.3 points worse at days 1 and 2 (exp 52; 3 wind farms; provisional).
CERRA runs about 3 months behind real time. It is deprioritised (data-sources page) (docs).

Blending experiments

Pass a shuffled-product placebo through every step, and give a guard arm the variant it guards (weather theme; exp 52, 54; blending arms; provisional).
Feed the run the service would have. Experiment 53 assumed a 3-hour UKV delivery time that no experiment has measured (exp 53; UKV; provisional).

Last checked against experiment 321.