rebase-energy/smhi-stations

SMHI Swedish Temperature Stations (Hourly)

Hourly air-temperature observations from 15 weather stations across Sweden, covering the last 10 years at 1-hour resolution. Built as a benchmark/evaluation dataset for autoregressive time-series forecasting.

• Source: SMHI Open Data — Meteorological Observations (metobs) API, parameter 1 (air temperature, momentanvärde, 1 gång/tim), corrected-archive (quality-controlled) period.
• Resolution: 1 hour (the finest available for SMHI temperature).
• Time zone: UTC.
• Unit: degrees Celsius (°C).
• Quality: only quality-controlled observations (codes G and Y) are kept.

Files

File	Description
swedish-temperatures.parquet	Main dataset. Wide layout: a regular hourly DatetimeIndex (UTC, named timestamp) × a 2-level MultiIndex of columns (station_id, station_name); values are temperature in °C. Gaps in the regular hourly grid are explicit NaN.
swedish-temperatures-stations.parquet	Station metadata: station_id, station_name, city, latitude, longitude, elevation (m), first_obs, last_obs, n_obs, pct_missing.

Stations

Stockholm, Göteborg, Malmö, Uppsala, Norrköping, Linköping, Örebro, Karlstad, Sundsvall, Östersund, Umeå, Luleå, Kiruna, Visby, Jönköping — geographically diverse, currently-active stations, each with < 5 % missing over the window.

Loading

The main file uses a pandas MultiIndex column layout, so load it with pandas:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="rebase-energy/smhi-stations",
    filename="swedish-temperatures.parquet",
    repo_type="dataset",
)
df = pd.read_parquet(path)                  # restores MultiIndex columns + UTC index
# df.columns.names == ['station_id', 'station_name']; df.index.name == 'timestamp'

# one station's series:
stockholm = df.xs("98230", level="station_id", axis=1)

Station metadata:

meta = pd.read_parquet(hf_hub_download(
    repo_id="rebase-energy/smhi-stations",
    filename="swedish-temperatures-stations.parquet",
    repo_type="dataset",
))

Suggested benchmark task

Autoregressive forecasting of hourly temperature per station. A simple persistence baseline (ŷ_{t+1} = y_t) yields a mean absolute error of ~0.66 °C across stations — a reasonable lower bar to beat. The regular hourly grid with explicit NaN gaps makes it straightforward to construct fixed-horizon train/eval splits.

License & attribution

The underlying observations are produced by SMHI (Swedish Meteorological and Hydrological Institute) and distributed as open data under Creative Commons Attribution 4.0 (CC BY 4.0). Please attribute SMHI when using this dataset. See https://opendata.smhi.se/ for the source and terms.

Reproducing

This dataset is built by scripts/build_swedish_temperatures.py in the emflow repository.