PEMS-BAY Traffic Dataset (TsFile)

This repository contains a TsFile conversion of witgaw/PEMS-BAY, a Hugging Face packaging of the PEMS-BAY traffic forecasting dataset derived from the DCRNN benchmark.

PEMS-BAY is used for traffic time-series forecasting, especially with graph neural networks such as Diffusion Convolutional Recurrent Neural Networks (DCRNN). The source dataset contains traffic flow sequences for 325 Bay Area traffic sensors, sampled at 5-minute intervals. Each sample keeps the chronological reference timestamp, 12 historical input steps, and 12 future target steps for a 1-hour prediction horizon.

Dataset Structure

The original dataset is distributed as Parquet files with chronological train/validation/test splits. This converted repository stores the same split structure as TsFile shards:

Split	Source split	TsFile pattern	TsFile files	Rows	Sensors	Time range
train	train.parquet	pems_bay_train_*.tsfile	48	11,851,125	325	2017-01-01 00:55:00 to 2017-05-07 16:35:00
validation	val.parquet	pems_bay_validation_*.tsfile	7	1,692,925	325	2017-05-07 16:40:00 to 2017-05-25 18:40:00
test	test.parquet	pems_bay_test_*.tsfile	14	3,386,175	325	2017-05-25 18:45:00 to 2017-06-30 22:55:00

Total converted rows: 16,930,225. The temporal split policy follows the source dataset: the earliest 70% of samples are used for training, the next 10% for validation, and the latest 20% for testing, preserving chronological order to avoid data leakage. All splits include all 325 sensors.

Schema

The source columns are mapped to TsFile table columns as follows:

Source column	TsFile role	Converted column
t0_timestamp	Time	Parsed as millisecond epoch time
node_id	TAG	node_id, converted to string while preserving values 0-324
x_t-11_d0 ... x_t+0_d0	FIELD	Historical traffic-flow values
x_t-11_d1 ... x_t+0_d1	FIELD	Historical normalized time-of-day features
y_t+1_d0 ... y_t+12_d0	FIELD	Future traffic-flow targets
y_t+1_d1 ... y_t+12_d1	FIELD	Future normalized time-of-day features

The TsFile table names follow the shard split: pems_bay_train, pems_bay_validation, and pems_bay_test.

Conversion Notes

• Only the converted .tsfile shards are hosted in this repository.
• The original Parquet files and the static sensor_graph/ files remain in the source dataset at witgaw/PEMS-BAY.
• t0_timestamp is encoded into the TsFile Time column and is not duplicated as a FIELD.
• node_id is stored as a TsFile TAG so each sensor can be queried as a device dimension.
• Source feature names containing + or - are normalized for TsFile schema compatibility, for example x_t-11_d0 becomes x_t_minus_11_d0 and y_t+12_d1 becomes y_t_plus_12_d1.
• The conversion preserves all 48 historical/input and future/target feature columns.

Usage

from tsfile import ColumnCategory, TsFileReader

reader = TsFileReader("pems_bay_train_1.tsfile")
schemas = reader.get_all_table_schemas()
table = "pems_bay_train"
columns = [
    column.get_column_name()
    for column in schemas[table].get_columns()
    if column.get_category() in (ColumnCategory.TAG, ColumnCategory.FIELD)
]

with reader.query_table(table, columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())

Source Dataset

• Original Hugging Face dataset: witgaw/PEMS-BAY
• Dataset generation code referenced by the source card: witgaw/DCRNN
• Original DCRNN repository/license: liyaguang/DCRNN

The source sensor graph metadata includes sensor coordinates, pairwise distances, a 325 x 325 precomputed adjacency matrix, and adjacency-generation metadata for graph neural network experiments.

Citation

If you use this dataset, please cite the original DCRNN paper:

@inproceedings{li2018dcrnn_traffic,
  title={{Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting}},
  author={{Li, Yaguang and Yu, Rose and Shahabi, Cyrus and Liu, Yan}},
  booktitle={{International Conference on Learning Representations}},
  year={{2018}}
}

License

This converted dataset follows the source dataset license: MIT.