Sub-daily FRP forecasting over Mediterranean wildfires: derived dataset (162 events)

Hourly fire radiative power for 162 Mediterranean wildfires, paired with co-located ERA5 meteorology and packaged as a ready-to-train forecasting benchmark. Each event is a burnt-area episode from the European Forest Fire Information System, bounded by its burn envelope plus a 0.25-degree buffer and followed from one day before the earliest fire date to nine days after the latest. Inside that box, 15-minute FRP retrievals from the EUMETSAT LSA SAF FRP-PIXEL product on MSG/SEVIRI are summed across every detected fire pixel and aggregated to clock hours; ERA5 fields are reduced to an area mean over the same box and hour. The result is one FRP series and one meteorological series per fire, on a common hourly grid. WHAT IS IN IT The events span 2007 to 2026 across 16 countries and are grouped by ignition year: - Development pool: 84 events, ignitions 2007 to 2024, split into 71 training and 13 validation.- 2025 test block: 46 events, every quality-screened fire of that season.- 2026 test block: 32 events, ignitions through 18 August 2026. All 162 pass the same fixed admission gate: at least 15 fire-active hourly slots, at least five days of duration, at least three diurnally complete days, a missing-slot rate below 40 percent and a saturation rate below 30 percent. No intensity threshold is applied, so peak hourly FRP ranges from about 0.9 to 150 GW. Files: - cube/dataset_v6_1.h5 — the model-ready cube (see the note below).- cube/dataset_v6_1_manifest.csv — event-to-split manifest and window accounting.- cube/normalization_params_v6_1.json — training-split normalization statistics.- hourly_frp/ — hourly area-aggregate FRP in MW, one file per event, 162 events.- hourly_era5/ — hourly ERA5 covariates, one file per event, 162 events.- metadata/ — catalogue, pools, splits and block lists. The cube holds supervised windows: a 48-hour lookback paired with a 12-hour horizon, advancing one hour at a time, over 32 channels. Four are temporal encodings, 14 are derived from the FRP series itself (lags, rolling statistics, trend, cumulative total, hours since first detection, a fire-active flag and the saturated-pixel fraction), and 14 come from ERA5 (winds at 10 m and 850 hPa, temperature, dewpoint, boundary-layer height, vapour pressure deficit, precipitation and the three-hour wind-speed change). metadata/event_catalogue_162.csv is the index: one row per event with its pool and split, ignition date, country, bounding box, the five gate metrics and whether each criterion passed. WHY THE TEST BLOCKS ARE NOT IN THE CUBE dataset_v6_1.h5 contains the training and validation groups only. The 78 test fires are deliberately absent and are assembled from the hourly tables at scoring time. That is not an oversight. The study these data support pre-registered a single evaluation on the 2025 block, and keeping those fires out of the cube is the mechanism that enforced it. Their hourly tables are here in full, so anyone reusing the data for a different question can build whatever splits they like. PROVENANCE, AND TWO CAVEATS WORTH KNOWING FRP comes from the EUMETSAT LSA SAF FRP-PIXEL product. Saturated pixels are kept at their reported values, which understate the true radiative power, and the saturated fraction travels with the data as a feature so it can be modelled or excluded. Meteorology comes from ERA5 for the 2025 block and for one February 2026 event, and from ERA5T, the preliminary release, for the other 31 events of the 2026 block, whose windows fall inside ERA5's finalisation lag of two to three months. ERA5T values can be revised, so results on the 2026 block may shift slightly against a later re-pull. Raw third-party products are not redistributed here. The acquisition pipeline that regenerates these tables from the original archives is published in the companion repository. REUSE Code, pre-registration, model checkpoints and every per-event result are at https://github.com/dmvalsamis/frp-prt-grsl (release v2.0.0). That repository includes a staging script that lays these files out where its analysis scripts expect them. This version supersedes the 34-event deposit (dataset_v3.h5) of an earlier version of the study, which remains available under its own version DOI. Licence CC BY 4.0. Attribution is also required by the upstream providers: the FRP data derive from the EUMETSAT LSA SAF FRP-PIXEL product, and the meteorological covariates from ERA5 and ERA5T, produced by the Copernicus Climate Change Service. Neither EUMETSAT nor C3S is responsible for any use of these derived products, and neither endorses this work.