TorchSpatiotemporal/ProcessedDatasets

Processed datasets for TSL

This repository contains the processed data files used by the Torch Spatiotemporal (TSL) dataset classes. The files are kept with the layout expected by those classes, so they can be downloaded, unpacked, and built by TSL without an additional conversion step.

The collection is intended primarily for reproducibility studies and benchmarks in TSL. It is not a new re-release of the underlying datasets: the datasets retain the provenance and usage conditions of their original sources. Please consult the original papers and providers before redistributing or using them.

Contents

Dataset	TSL class	Data	Processed size
PeMS03	tsl.datasets.PeMS03	Traffic flow from 358 California sensors, sampled every 5 minutes from September–November 2018.	26,208 × 358 × 1
PeMS04	tsl.datasets.PeMS04	Traffic flow, occupancy, and speed from 307 San Francisco Bay Area sensors, sampled every 5 minutes from January–February 2018.	16,992 × 307 × 3
PeMS07	tsl.datasets.PeMS07	Traffic flow from 883 California sensors, sampled every 5 minutes from May–August 2017.	28,224 × 883 × 1
PeMS08	tsl.datasets.PeMS08	Traffic flow, occupancy, and speed from 170 San Bernardino sensors, sampled every 5 minutes from July–August 2016.	17,856 × 170 × 3
PeMS-Bay	tsl.datasets.PemsBay	Traffic readings from 325 San Francisco Bay Area sensors, sampled every 5 minutes from January–June 2017.	52,128 × 325 × 1
METR-LA	tsl.datasets.MetrLA	Traffic readings from 207 Los Angeles loop detectors, sampled every 5 minutes from March–June 2012.	34,272 × 207 × 1
Air Quality	tsl.datasets.AirQuality	PM2.5 measurements from 437 stations in 43 Chinese cities, sampled hourly from May 2014–April 2015. Includes the 36-station Beijing subset.	8,760 × 437 × 1, or 8,760 × 36 × 1
LargeST	tsl.datasets.LargeST	Five years of California traffic readings from 8,600 sensors, sampled every 5 minutes from 2017–2021. TSL can select CA, GLA, GBA, or SD subsets and individual years.	525,888 × 8,600 × 1
PV-US	tsl.datasets.PvUS	Synthetic photovoltaic power production for 5,016 US plants, sampled every 10 minutes. East and west zones are provided separately.	52,560 × 5,016 × 1

Dimensions above describe the main time-varying target. Some datasets also include masks, distances, locations, metadata, occupancy, speed, or adjacency information.

File layout

The layout is part of the interface with TSL:

air/
  air_quality.zip               # full437.h5 and small36.h5
traffic/
  metr_la.zip                   # METR-LA data and sensor metadata
  pems03.zip                    # PeMS03 data and distances
  pems04.zip
  pems07.zip
  pems08.zip
  pems_bay.zip                  # PeMS-Bay data and sensor metadata
large_st/
  2017/data.h5 ... 2021/data.h5
  sensors.zip                   # adj.npz and metadata.csv
pv_us/
  east.h5
  west.h5

The ZIP files should remain archives. They contain the raw files that the corresponding TSL class expects when it runs its download()/build() pipeline. LargeST is split by year so callers can request only the years they need.

Using the data with TSL

After installing TSL, extract the relevant archive into a loader root (or arrange the uncompressed files in the layout expected by the class), then pass that root to the class:

from tsl.datasets import AirQuality, LargeST, MetrLA, PeMS04, PvUS

air = AirQuality(root="/path/to/extracted/air_quality")
traffic = PeMS04(root="/path/to/extracted/pems04")
metr = MetrLA(root="/path/to/extracted/metr_la")
large = LargeST(root="/path/to/ProcessedDatasets/large_st", year=2019)
solar = PvUS(root="/path/to/ProcessedDatasets/pv_us")

For an automatic download through TSL, use a TSL version whose dataset URLs point to this Hugging Face repository. The repository preserves the filenames needed by the loaders; the URL migration itself belongs in the TSL dataset class definitions.

Sources and related work

• PeMS03/04/07/08: CalTrans PeMS data, distributed in the benchmark setup of Guo et al. (2021).
• METR-LA and PeMS-Bay: the benchmark data used by Li et al. (2018).
• Air Quality: the Urban Air project at Microsoft Research.
• LargeST: Liu et al. (2023), “LargeST: A Benchmark Dataset for Large-Scale Traffic Forecasting”.
• PV-US: synthetic photovoltaic data from NREL’s Solar Power Data for Integration Studies.

See the docstrings of the corresponding classes in TSL for preprocessing details, missing-value handling, covariates, graph construction, and dataset-specific splitters.

License and responsible use

The original datasets have different providers and terms. No single license is asserted here for all files. Users are responsible for complying with the terms of the original sources and for citing the relevant source papers. The files are supplied for research, education, reproducibility, and benchmarking; they should not be treated as real-time operational data.