Gap-filled wind and irradiance records for German weather stations, 2023-2026

Wind and solar resource assessment, forecast verification and the validation of gridded products all start from measured station data — and station data has holes. Sensors ice up, pyranometer domes soil, loggers fail, masts go down for maintenance. In this German network, 3260 interior gaps interrupt three years of record, and the gaps longer than two days hold more than three quarters of the missing time. Temporal interpolation, the state of practice, draws on nothing but the series it is trying to repair, so it cannot bridge them. This dataset closes those gaps. It contains **hourly wind speed and direction at 204 stations** and **half-hourly global and diffuse horizontal irradiance at 94 stations** of the German Meteorological Service (Deutscher Wetterdienst), covering **2023-07-24 to 2026-07-31** in UTC, with every gap filled and **every value flagged as measured or filled**. What is in it - **wind/** — 204 Parquet files, one per station, 5.4 million rows, no missing timestamps - **irradiance/** — 94 Parquet files, 5.0 million rows - **stations.csv** — coordinates, elevation, and how much was filled per station and variable - **README.md** — full column documentation Files are named by station identifier and station name. Every row carries a boolean flag per variable, named after it with the suffix **_imputed**: true where the value came from the model, false where it is the original measurement. Filled in total: 71 091 hours of wind speed, 71 616 hours of wind direction, and 275 thousand half-hourly steps of global and 286 thousand of diffuse irradiance. How the gaps were filled A temporal fusion transformer converted from forecasting to masked imputation, one model per resource for the whole network, with the station identity as a learned embedding. The model reads a gap-free meteorological source across the entire window — the ERA5 reanalysis for wind, and for irradiance the CAMS radiation service at the station and at eight surrounding points. Gaps longer than the window are filled by tiling the window across them. On held-out cells the wind model reaches a skill of 0.60 against ERA5 in a one-hour gap and 0.46 where no measurement is visible anywhere in the window, raising R² from 0.59 to 0.94 and 0.88. Against the satellite the irradiance skill is 0.29 and 0.24 for the global and 0.45 and 0.36 for the diffuse component. Skill is one minus the ratio of the model's root mean square error to that of the uncorrected source. One caveat worth reading Some stations joined the network during the period or stopped reporting before it ended, so part of what is filled lies before a station's first or after its last measurement. No measurement exists there, so those values could never be checked against one, and the evaluation above does not cover them. The README explains how to separate them from the interior gaps in four lines of code. Users who need values backed by the reported skill should restrict themselves to interior gaps. Attribution and licence This dataset is derived from third-party data, and their attribution terms apply in addition to the licence below. **Measurements.** Datenbasis: Deutscher Wetterdienst, values aggregated from ten-minute observations and gaps filled by a model. DWD open data is published under CC BY 4.0. **Covariates.** Generated using Copernicus Climate Change Service information 2026 and Copernicus Atmosphere Monitoring Service information 2026. Neither the European Commission nor ECMWF is responsible for any use that may be made of the Copernicus information or data it contains. **This dataset** is released under CC BY 4.0. Please keep the imputed flags with the data when redistributing, so that filled values remain distinguishable from measured ones. Method, evaluation and limitations are described in the accompanying paper.