smilegeng/ETDataset-ett

ETT (Electricity Transformer Temperature) — TsFile format

This repository is a conversion of the ETT (Electricity Transformer Temperature) dataset to Apache TsFile format.

• Original dataset (Hugging Face): ETDataset/ett
• Original data source (GitHub): zhouhaoyi/ETDataset
• Paper: Haoyi Zhou et al., Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting, AAAI 2021. arXiv:2012.07436
• License: CC-BY-4.0

Dataset Description

ETT records two years of electricity-transformer operating data from two regions in China and is a common benchmark for long-sequence time-series forecasting (LSTF). Each data point contains the target oil temperature (OT) and 6 power-load features. The original data is provided at two sampling rates (hourly / every 15 minutes), in 4 sequence files:

File	Region	Sampling	Rows	Time span
data/ETTh1.tsfile	Region 1	hourly	17,420	2016-07-01 ~ 2018-06-26
data/ETTh2.tsfile	Region 2	hourly	17,420	2016-07-01 ~ 2018-06-26
data/ETTm1.tsfile	Region 1	15-min	69,680	2016-07-01 ~ 2018-06-26
data/ETTm2.tsfile	Region 2	15-min	69,680	2016-07-01 ~ 2018-06-26

Column meanings

Column	Meaning	Type
Time	Timestamp (INT64, millisecond precision)	time column
HUFL	High UseFul Load	FLOAT
HULL	High UseLess Load	FLOAT
MUFL	Middle UseFul Load	FLOAT
MULL	Middle UseLess Load	FLOAT
LUFL	Low UseFul Load	FLOAT
LULL	Low UseLess Load	FLOAT
OT	Oil Temperature (forecast target)	FLOAT

Repository structure

ETDataset-ett/
├── README.md          # this file
└── data/
    ├── ETTh1.tsfile
    ├── ETTh2.tsfile
    ├── ETTm1.tsfile
    └── ETTm2.tsfile

Conversion Notes

• Data source: converted from the 4 original CSVs in the GitHub repo zhouhaoyi/ETDataset (the Hugging Face ETDataset/ett is a Python loading script that fetches these CSVs at runtime and reshapes them into GluonTS-style sliding-window samples; this repository converts the underlying original CSVs, not the reshaped train/val/test sliding-window view — that is just an index view over the same continuous series).
• Each CSV is converted to its own .tsfile, not merged; the full two-year series is kept with no train/val/test split.
• Time column: parsed from the original date string column (YYYY-MM-DD HH:MM:SS) into INT64 millisecond timestamps. The original date string column is not kept separately — its information is losslessly folded into the Time column.
• Field columns: HUFL/HULL/MUFL/MULL/LUFL/LULL/OT, 7 columns in total, all stored as single-precision FLOAT.
• No data columns dropped: date → Time (lossless); all 7 numeric columns retained.
• No TAG columns: each file is an independent time series, so no device dimension is needed.

Reading example

from tsfile import TsFileReader

reader = TsFileReader("data/ETTh1.tsfile")
schemas = reader.get_all_table_schemas()
tname = next(iter(schemas))
field_cols = [c.get_column_name() for c in schemas[tname].get_columns()]
with reader.query_table(tname, field_cols, batch_size=65536) as rs:
    while (batch := rs.read_arrow_batch()) is not None:
        df = batch.to_pandas()
        print(df.head())
        break

Citation

@inproceedings{haoyietal-informer-2021,
  author    = {Haoyi Zhou and Shanghang Zhang and Jieqi Peng and Shuai Zhang and
               Jianxin Li and Hui Xiong and Wancai Zhang},
  title     = {Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting},
  booktitle = {The Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021},
  volume    = {35},
  number    = {12},
  pages     = {11106--11115},
  publisher = {AAAI Press},
  year      = {2021}
}