Dataset Card — FaroeWRF-500

A 500 m WRF + observation benchmark for 1-hour-ahead weather forecasting over the Faroe Islands.

Summary

Field	Value
Name	FaroeWRF-500
File	faroe_wrf500_benchmark.h5
Version	1.0.1
Format	HDF5, 16 GB
License	CC-BY-4.0
Period	2025-03-01 — 2026-03-31 (397 days, 8867 hourly timesteps)
Grid	300 × 200 cells, 500 m horizontal resolution, WRF d02
Stations	28 (23 Landsverk RWIS + 5 Vedur AWS)
Cadences	hourly (all 28 stations), 5-minute (all 28), 1-minute (5 Vedur)
Host	https://huggingface.co/datasets/anon676767/faroewrf-500 (anonymous account during double-blind review; will be transferred to the authors' account at camera-ready)

Intended use

Designed to evaluate machine-learning weather emulators on heterogeneous coastal terrain where receptive field, Fourier-mode coverage, and filter locality measurably affect performance. The matched grid/station structure supports terrain-stratified evaluation (tier 1/2/3 by topographic representativeness, exposure class, and elevation mismatch with WRF terrain), sector-conditioned analysis (8 wind directions), and storm-intensity stratification.

Not intended for: operational nowcasting (no real-time pipeline), forecast verification of WRF itself (the WRF runs are inputs to the benchmark, not the object of study), or lead times beyond 1 hour without re-deriving targets.

Sources

• WRF d02 outputs. Operational forecast runs from the Faroese Meteorological Institute (Veðurstovan). Daily 00 Z initialisation, 25 hourly outputs, concatenated into 8867 hourly timesteps.
• Landsverk RWIS observations. Road-weather station observations published by Landsverk (Faroese roads administration). Native cadence 5 minutes; 5-min and hourly aggregates included in the file.
• Vedur AWS observations. Faroese Meteorological Institute hourly archive; native 1-minute cadence is included for the 5 AWS stations.

Variables

WRF (10 channels at both /wrf_grid (300 × 200) and /wrf_stations (28)): U10, V10, T2, Q2, PSFC, PBLH, HFX, LH, SWDOWN, GLW.

Hourly observations (28 stations): wind_speed, wind_direction, wind_gust, air_temperature, air_pressure (QFE), air_pressure_qff, relative_humidity, precipitation.

5-minute observations (28 stations) and 1-minute observations (5 Vedur AWS) follow analogous schemas; see BENCHMARK_DATASET_README.md for the full listing.

Unit convention. WRF temperature and pressure are in K and Pa; station observations are in °C and hPa. Convert before computing residuals.

See BENCHMARK_DATASET_README.md for the complete data dictionary, HDF5 group structure, and per-station metadata schema.

Quality control

• Range and spike detection on Landsverk and Vedur raw streams.
• Fatal flags masked to NaN; informational flags preserved upstream of the HDF5 file.
• Wind direction is a circular (vector) mean.
• Wind gust is the within-hour maximum; note WRF has no analogous sub-grid peak, so WRF "gust" verifications will systematically underread.

Known data issues are documented in BENCHMARK_DATASET_README.md (Mykines lighthouse anemometer height; Hogareyn pressure anomaly; Landsverk wind direction cadence; WRF orography smoothing).

Splits

The reproducibility scripts in scripts/compute_splits.py produce a deterministic temporal split (train / val / test) over the 397 daily cycles. A splits.json is written to data/processed/ and consumed by all training scripts. The split is part of the reproducibility artifact (it is not stored inside the HDF5).

Loading example

import h5py
import numpy as np

with h5py.File("faroe_wrf500_benchmark.h5", "r") as f:
    u10_grid = f["wrf_grid/U10"][:]              # (8867, 300, 200)
    u10_stn  = f["wrf_stations/U10"][:]          # (8867, 28)
    obs_wind = f["observations/wind_speed"][:]   # (8867, 28)
    ts = f["timestamps"][:].astype("datetime64[h]")
    stations = [s.decode() for s in f["metadata/stations"][:]]

Reviewer sample (1.25 GB)

A 30-day sample (faroe_wrf500_benchmark_sample.h5, 2025-12-01 to 2025-12-30) is provided so reviewers can inspect the data and exercise the full pipeline without downloading the 16 GB archive. The schema is byte-identical to the full file — the same loader code works on both.

The accompanying code artifact reads the env var FAROEWRF_HDF5 to pick which file to use. To run any script (or ./reproduce.sh) against the sample instead of the full archive:

export FAROEWRF_HDF5=data/benchmark/faroe_wrf500_benchmark_sample.h5
./reproduce.sh --quick      # for example

When the variable is unset, the code defaults to the full archive at data/benchmark/faroe_wrf500_benchmark.h5.

License

This dataset is released under Creative Commons Attribution 4.0 International (CC-BY-4.0).

You are free to share and adapt the dataset for any purpose, including commercially, provided you give appropriate credit and indicate any changes made.

When redistributing or building on this dataset, please attribute the original sources:

• WRF and AWS data: Veðurstovan (Faroese Meteorological Institute)
• RWIS data: Landsverk (Faroese roads administration)

Citation

@dataset{anonymous2026faroewrf500,
  title     = {FaroeWRF-500: A High-Resolution Weather Benchmark for the Faroe Islands},
  author    = {Anonymous Authors},
  year      = {2026},
  publisher = {Hugging Face},
  version   = {1.0.1},
  url       = {https://huggingface.co/datasets/anon676767/faroewrf-500},
  note      = {NeurIPS 2026 Datasets and Benchmarks Track; author list anonymised under review}
}

Maintenance and contact

• During double-blind review the dataset is hosted under an anonymous Hugging Face account; the OpenReview submission contains the dataset URL.
• Bug reports / errata: please use the OpenReview discussion thread during review. After acceptance, the repository will be transferred to the authors' Hugging Face account and a discussion thread on the dataset card will become the primary contact channel.
• Versioning: HF Datasets git-style revisions are used; main resolves to the latest version, and tagged revisions preserve specific releases (e.g. v1.0.0, v1.0.1).

Changelog

• 1.0.1 (2026-05-04): Per-station metadata refreshed against the 2020 Landsverk + DMI/Vedur on-site surveys: GNSS-measured lat/lon, levelled elevation, and measured wind_height_m, barometer_height_m, temp_height_m. 25 of 28 stations now carry an authoritative anemometer height (was 5 of 28 in v1.0.0; the three remaining stations had no wind sensor at survey time). delta_h recomputed against the existing elevation_wrf. Bulk WRF and observation arrays are unchanged.
• 1.0.0 (initial public release, NeurIPS 2026): 13-month archive (2025-03-01 to 2026-03-31, 8867 hourly timesteps), 28 stations, full WRF d02 grid + station-interpolated fields, plus 5-minute and 1-minute observation subgroups.