weatherAUS — Australian Daily Weather (TsFile)

Daily weather observations from 49 locations across Australia, the classic "predict if it rains tomorrow" dataset, converted from Parquet to Apache TsFile format.

Each time series is one weather station/city, observed once per day from 2007-11-01 to 2017-06-25, across 22 meteorological variables (temperature, rainfall, humidity, pressure, wind, cloud, sunshine, …).

Source

• Original dataset: lowkeydev-ln/weatherAUS
• Underlying data: Australian Bureau of Meteorology (BOM) public weather observations. The dataset is widely circulated (e.g. on Kaggle) as "Rain in Australia".
• License: the source HF dataset card declares no explicit license; the underlying observations are public BOM data. Verify terms before redistribution.

Data structure

Single Parquet (142,193 rows × 24 columns), long format: one row per (Location, Date). Key facts derived from the data:

• 49 distinct locations (Location, e.g. Adelaide, Sydney, Woomera), each an independent daily time series (≈1,521–3,418 rows each).
• Date is a real calendar date (YYYY-MM-DD); daily, equally spaced.
• (Location, Date) is unique — no duplicate (station, day) rows.
• Many columns contain missing values (e.g. Sunshine ~67k, Cloud3pm ~57k, Evaporation ~61k missing). These are preserved as-is — no imputation, no rows dropped.

TsFile mapping

The TsFile uses the table model. One location = one device (time series).

TsFile role	Column(s)	Type	Notes
TAG (device)	Location	STRING	49 Australian cities/stations
TIME	derived from Date	INT64 (ms)	real Unix-epoch milliseconds
FIELD	MinTemp, MaxTemp, Rainfall, Evaporation, Sunshine, WindGustSpeed, WindSpeed9am, WindSpeed3pm, Humidity9am, Humidity3pm, Pressure9am, Pressure3pm, Cloud9am, Cloud3pm, Temp9am, Temp3pm, RISK_MM	DOUBLE	numeric measurements
FIELD	WindGustDir, WindDir9am, WindDir3pm	STRING	16-point compass direction
FIELD	RainToday, RainTomorrow	STRING	Yes / No

• Table name: weather_aus. Time precision: ms.
• Within each device, rows are sorted by Time (ascending).

Conversion notes (column handling)

• Date → Time: parsed into the derived Time column (Unix epoch ms) and not kept as a separate field. Lossless — daily, equally spaced.
• RISK_MM kept: this is the well-known target-leakage column (it equals the next day's Rainfall, i.e. the answer behind RainTomorrow). It is kept here as a genuine measurement; drop it if you train a model on this data.
• No rows dropped, no imputation. Missing numeric values become NaN, missing strings become empty.
• Single split (no separate train/test); the entire CSV is one TsFile.

Files

.
├── README.md
└── data/
    └── weather_aus.tsfile

Reading the TsFile

from tsfile import TsFileReader

reader = TsFileReader("data/weather_aus.tsfile")
for name, table in reader.get_all_table_schemas().items():
    print(name, [c.get_column_name() for c in table.get_columns()])
# Query FIELD/TAG columns of the table 'weather_aus' via reader.query_table(...)