akderekaan/de-power-forecast-data

DE Power Forecast – data

Input and target data for forecasting German solar, onshore wind and offshore wind generation, collected automatically by the GitHub Actions jobs in ak16o7/de-power-forecast.

The rule behind the layout: every value carries the time it became available. A forecast issued at time T may only use rows with seen_at / available_at ≤ T. That is what makes backtests on this data honest.

Files

Path	What	Time columns
entsoe/a75/{area}/{YYYY}/{YYYY-MM}.parquet	Actual generation (ENTSO-E A75), 15 min, MW, latest known value	ts (interval start, UTC), fetched_at
entsoe/a69_da/…, entsoe/a69_id/…	TSO wind & solar forecasts, day-ahead (A01, published ~18:00 CET D-1) and intraday (A40, ~08:00 CET D)	ts, fetched_at
entsoe/a69_current/…	TSO continuously updated forecast (A18), final archived version only: revised until ~30–80 min after delivery starts, so not a vintage (see vintages/ for versions as seen)	ts, fetched_at
vintages/entsoe/{YYYY}/{YYYY-MM-DD}.parquet	Recorder log every 15 min: each new or changed value of A75 and A69 (day-ahead, intraday, current) with the time we saw it	ts, seen_at
vintages/entsoe/parts/{day}/…	Same for the current UTC day, compacted after midnight
weather/runs/{model}/{YYYY-MM}/{run}.parquet	One file per weather model run, 16 points, hourly	run, valid, available_at, fetched_at, source (live = recorded when published, archive = backfilled)
weather/previous_runs/{model}/{YYYY}/{YYYY-MM}.parquet	Values predicted ≥ 24 h (lead_days=1) and ≥ 48 h (lead_days=2) before valid	valid, available_at = valid − N·24 h + model delay (upper bound)
capacity/latest.parquet, capacity/snapshots/{date}.parquet	Installed capacity per month (GW), snapshots because it is revised later	month, seen_at
reports/baseline.json	Baseline scores: TSO forecasts, persistence, naive day-ahead rules; MAE, RMSE, bias, nMAE per technology and lead time, overall, last 12 months and per month	generated_at
reports/model_da.json, reports/model_da_backtest.parquet	Day-ahead model v1: walk-forward backtest (P10/P50/P90 per quarter hour) and scores against the TSO day-ahead forecast	ts, generated_at
reports/model_id.json, reports/model_id_backtest.parquet	Intraday model v2 (correction of the newest published TSO forecast, 15 min to 8 h): walk-forward scores against TSO, TSO plus last error and persistence	generated_at
state/*.json, status.json	Job bookkeeping and health

Areas: DE (Germany) and the control areas 50HZ, AMP (Amprion), TTG (TenneT), TBW (TransnetBW). Targets: solar, wind_on, wind_off.

Weather models: ecmwf_ifs (ECMWF IFS HRES 9 km), icon_d2 (DWD ICON-D2 2 km), icon_eu (DWD ICON-EU 7 km). Points: 13 onshore regions and 3 offshore sites (North Sea ×2, Baltic Sea), see dpf/config.py in the code repo.

Sources and licences

• ENTSO-E Transparency Platform (A75, A69) – data item on ENTSO-E's list of data for free re-use, CC BY 4.0. Source: ENTSO-E Transparency Platform.
• Weather: Open-Meteo (CC BY 4.0), underlying models by ECMWF and Deutscher Wetterdienst.
• Installed capacity: Energy-Charts (Fraunhofer ISE), CC BY 4.0.

This dataset is published under CC BY 4.0; please credit the sources above. Values are provided as received, without warranty. Actuals are provisional until the TSOs publish metered values.