smilegeng/dmi-aarhus-weather-data

DMI Aarhus Weather Data (TsFile)

This is the Apache TsFile conversion of the time-series tables from Ciroc0/dmi-aarhus-weather-data.

Training data and observation context for the Aarhus (Denmark) weather forecast-correction pipeline. The original repository also ships several trained model artifacts (*.pkl, model_registry.json, model_meta.json); those are not time-series data and are not included here — only the two Parquet tables are converted.

• Location: Aarhus, Denmark — 56.1567, 10.2108, timezone Europe/Copenhagen
• Upstream data: DMI / DMI HARMONIE forecasts and Open-Meteo observations.

Files

.
├── README.md
└── data/
    ├── data.tsfile               (339,773 bytes)     — from data.parquet (legacy)
    └── training-matrix.tsfile    (31,262,352 bytes)  — from training_matrix.parquet (current)

All timestamps from the source (timezone Europe/Copenhagen) are converted to UTC epoch milliseconds for the Time axis — the physical instant is unchanged, only the stored representation is timezone-normalized.

training-matrix.tsfile — current source of truth

From training_matrix.parquet (41 528 rows × 76 columns). Each target hour is forecast multiple times from different forecast-issue times, so a single target_timestamp is not unique on its own. In TsFile this is modeled as one table with many devices, where the device tag is the forecast-issue time.

• Table: dmi_aarhus_training_matrix
• TAG (device): reference_time — the forecast issue time rendered as a UTC ISO-8601 string (e.g. 2026-01-20T23:00:00Z). One device per forecast-issue time (873 devices).
• Time: target_timestamp → UTC epoch ms. Unique within each device.
• FIELD: all remaining 74 columns, grouped below.

group	columns	meaning
forecast horizon	lead_time_hours (INT64), lead_bucket (STRING: 1-6/7-12/13-24/25-48)	hours ahead of the issue time, bucketed
location	latitude, longitude	constant (Aarhus)
DMI forecast (dmi_*_pred)	temperature, apparent temp, humidity, dew point, pressure, cloud cover (total/low/mid/high), precipitation, rain, snowfall, precip prob, wind speed/direction/gusts, visibility, radiation (shortwave/direct), weather code, CAPE	raw DMI HARMONIE forecast values for the target hour
forecast wind components	forecast_wind_u, forecast_wind_v	decomposed forecast wind vector
run deltas (*_run_delta)	temperature, wind speed, wind gusts, precipitation, pressure, relative humidity	change vs. the previous model run
actuals (actual_*)	temp, humidity, pressure, precipitation, rain, wind speed/direction/gust, wind u/v	observed values at the target hour
rain targets	rain_event (INT64 0/1), rain_amount	observed rain occurrence / amount
calendar	hour, month, day_of_year, hour_sin, hour_cos, month_sin, month_cos	cyclic time features
observation context	observation_context_timestamp (INT64 UTC epoch ms), obs_* lag / rolling-mean / rolling-sum features (temp, wind, wind u/v, pressure, humidity, precip)	causal history features available at issue time
correction targets	temp_correction_target, wind_speed_correction_target, wind_gust_correction_target	supervised regression targets (actual − forecast); leading nulls where no DMI forecast exists for early issue times are kept as null

observation_context_timestamp is a timezone-aware timestamp in the source; it is stored as an INT64 UTC epoch-milliseconds field (a plain numeric field, not the device time axis and not a string).

data.tsfile — legacy compatibility table

From data.parquet (2 382 rows × 20 columns). A single clean series — the timestamp column is fully unique — so it is stored as one device, no tag.

• Table: dmi_aarhus_data
• Time: timestamp → UTC epoch ms (unique)
• FIELD: reference_time (INT64 UTC epoch ms), lead_time_hours, dmi_temp_pred, dmi_wind_pred, dmi_pressure_pred, dmi_humidity_pred, actual_temp, actual_wind, actual_pressure, actual_humidity, hour, day_of_year, month, hour_sin, hour_cos, month_sin, month_cos, dmi_error
• Dropped: __index_level_0__ (a leftover pandas RangeIndex — 0,1,2,…, no information).

This table is marked Legacy upstream (older single-target layout); the training-matrix.tsfile is the current multi-target source of truth.

Reading the TsFiles

from tsfile import TsFileReader

reader = TsFileReader("data/training-matrix.tsfile")
for name, table in reader.get_all_table_schemas().items():
    print(name, [(c.get_column_name(), c.get_data_type()) for c in table.get_columns()])

cols = ["lead_time_hours", "dmi_temperature_2m_pred",
        "actual_temp", "temp_correction_target"]
with reader.query_table("dmi_aarhus_training_matrix", cols, batch_size=65536) as rs:
    while (batch := rs.read_arrow_batch()) is not None:
        df = batch.to_pandas()   # Time + reference_time (tag) added automatically
        ...

Note on querying tags. In the TsFile table model, a TAG column (reference_time here) is a device-identifying dimension, not an independently queryable measurement. A query must include at least one FIELD column; the TAG and Time columns are then returned alongside it. Querying only the tag column returns 0 rows — this is expected table-model behavior, not a corrupt file.

Conversion notes

• All Time columns are UTC epoch milliseconds; source timezone (Europe/Copenhagen) is normalized to UTC, physical instants unchanged.
• The training matrix is keyed by (reference_time, target_timestamp) in the source (each target hour is forecast once per issue time). It is modeled with reference_time as the device TAG and target_timestamp as the Time axis, so the time axis is naturally unique within each device — no timestamps are altered or offset.
• Timezone-aware non-axis timestamps (reference_time in the legacy data table, observation_context_timestamp in the training matrix) are stored as INT64 UTC epoch-millisecond fields (plain numeric fields).
• Missing values are preserved as null (e.g. the dmi_*_pred and *_correction_target columns are null for early issue times that predate the DMI forecast feed); nothing is imputed or dropped.
• The .pkl model bundles and *.json registries from the source repo are not time-series data and are not converted here.

Source & license

Released under CC BY 4.0. Please preserve attribution to:

• Ciroc0 — original dataset author

• Open-Meteo — observation data

• DMI / DMI HARMONIE — forecast data

• Original Hugging Face dataset: Ciroc0/dmi-aarhus-weather-data

Upstream attribution requirements from the data providers still apply.