metadata

license: apache-2.0
configs:
  - config_name: default
    data_files:
      - split: train
        path: '*/*.arrow'
  - config_name: UTSD-1G
    data_files:
      - split: train
        path: UTSD-1G/*.arrow
  - config_name: UTSD-2G
    data_files:
      - split: train
        path: UTSD-2G/*.arrow
  - config_name: UTSD-4G
    data_files:
      - split: train
        path: UTSD-4G/*.arrow
  - config_name: UTSD-12G
    data_files:
      - split: train
        path: UTSD-12G/*.arrow
task_categories:
  - time-series-forecasting
tags:
  - time series forecasting
  - time series analysis
  - time series
pretty_name: UTSD
size_categories:
  - 100M<n<1B

Unified Time Series Dataset (UTSD)

Updates

:triangular_flag_on_post: News (2024.10) We release the numpy format of UTSD. An easier and more efficient dataloader can be found here.

Introduction

We curate Unified Time Series Dataset (UTSD) that includes 7 domains with up to 1 billion time points with hierarchical four volumes to facilitate research of large models and pre-training in the field of time series.

Unified Time Series Dataset (UTSD) is meticulously assembled from a blend of publicly accessible online data repositories and empirical data derived from real-world machine operations.

All datasets are classified into seven distinct domains by their source: Energy, Environment, Health, Internet of Things (IoT), Nature, Transportation, and Web with diverse sampling frequencies.

See the paper and codebase for more information.

Dataset detailed descriptions.

We analyze each dataset, examining the time series through the lenses of stationarity and forecastability to allow us to characterize the level of complexity inherent to each dataset.

Domain	Dataset	Time Points	File Size	Freq.	ADF.	Forecast.	Source
Energy	London Smart Meters	166.50M	4120M	Hourly	-13.158	0.173	[1]
Energy	Wind Farms	7.40M	179M	4 sec	-29.174	0.811	[1]
Energy	Aus. Electricity Demand	1.16M	35M	30 min	-27.554	0.730	[1]
Environment	AustraliaRainfall	11.54M	54M	Hourly	-150.10	0.458	[2]
Environment	BeijingPM25Quality	3.66M	26M	Hourly	-31.415	0.404	[2]
Environment	BenzeneConcentration	16.34M	206M	Hourly	-65.187	0.526	[2]
Health	MotorImagery	72.58M	514M	0.001 sec	-3.132	0.449	[3]
Health	SelfRegulationSCP1	3.02M	18M	0.004 sec	-3.191	0.504	[3]
Health	SelfRegulationSCP2	3.06M	18M	0.004 sec	-2.715	0.481	[3]
Health	AtrialFibrillation	0.04M	1M	0.008 sec	-7.061	0.167	[3]
Health	PigArtPressure	0.62M	7M	-	-7.649	0.739	[3]
Health	PigCVP	0.62M	7M	-	-4.855	0.577	[3]
Health	IEEEPPG	15.48M	136M	0.008 sec	-7.725	0.380	[2]
Health	BIDMC32HR	63.59M	651M	-	-14.135	0.523	[2]
Health	TDBrain	72.30M	1333M	0.002 sec	-3.167	0.967	[5]
IoT	SensorData	165.4M	2067M	0.02 sec	-15.892	0.917	Real-world machine logs
Nature	Phoneme	2.16M	25M	-	-8.506	0.243	[3]
Nature	EigenWorms	27.95M	252M	-	-12.201	0.393	[3]
Nature	ERA5 Surface	58.44M	574M	3 h	-28.263	0.493	[4]
Nature	ERA5 Pressure	116.88M	1083M	3h	-22.001	0.853	[4]
Nature	Temperature Rain	23.25M	109M	Daily	-10.952	0.133	[1]
Nature	StarLightCurves	9.46M	109M	-	-1.891	0.555	[3]
Nature	Saugen River Flow	0.02M	1M	Daily	-19.305	0.300	[1]
Nature	KDD Cup 2018	2.94M	67M	Hourly	-10.107	0.362	[1]
Nature	US Births	0.00M	1M	Daily	-3.352	0.675	[1]
Nature	Sunspot	0.07M	2M	Daily	-7.866	0.287	[1]
Nature	Worms	0.23M	4M	0.033 sec	-3.851	0.395	[3]
Transport	Pedestrian Counts	3.13M	72M	Hourly	-23.462	0.297	[1]
Web	Web Traffic	116.49M	388M	Daily	-8.272	0.299	[1]

You can find the specific source address in source.csv.

[1]: Monash Time Series Forecasting Archive [2]: Time series extrinsic regression Predicting numeric values from time series data [3]: The UCR Time Series Archive [4]: ERA5-Land: a state-of-the-art global reanalysis dataset for land applications [5]: Contrast Everything: A Hierarchical Contrastive Framework for Medical Time-Series

Hierarchy of Datasets

UTSD is constructed with hierarchical capacities, namely UTSD-1G, UTSD-2G, UTSD-4G, and UTSD-12G, where each smaller dataset is a subset of the larger ones. A larger subset means greater data difficulty and diversity, allowing you to conduct detailed scaling experiments.

Usage

You can access and load UTSD based on the code in our repo.

# huggingface-cli login
# export HF_ENDPOINT=https://hf-mirror.com 

python ./scripts/UTSD/download_dataset.py

# dataloader
python ./scripts/UTSD/utsdataset.py

It should be noted that due to the construction of our dataset with diverse lengths, the sequence lengths of different samples vary. You can construct the data organization logic according to your own needs.

In addition, we provide code dataset_evaluation.py for evaluating time series datasets, which you can use to evaluate your Huggingface formatted dataset. The usage of this script is as follows:

python dataset_evaluation.py --root_path <dataset root path> --log_path <output log path>

Acknowledgments

UTSD is mostly built from the Internet public time series dataset, which comes from different research teams and providers. We sincerely thank all individuals and organizations who have contributed the data. Without their generous sharing, this dataset would not have existed.

Citation

If you're using UTSD in your research or applications, please cite it using this BibTeX:

BibTeX:

@inproceedings{liutimer,
  title={Timer: Generative Pre-trained Transformers Are Large Time Series Models},
  author={Liu, Yong and Zhang, Haoran and Li, Chenyu and Huang, Xiangdong and Wang, Jianmin and Long, Mingsheng},
  booktitle={Forty-first International Conference on Machine Learning}
}