BTS: Building Timeseries Dataset

Anton Grabolle / Better Images of AI / AI Architecture / Licenced by CC-BY 4.0

BTS: Building Timeseries Dataset: Empowering Large-Scale Building Analytics.

[ArXiv] [OpenReview] [NeurIPS] [Slides] [Poster]

The Building TimeSeries (BTS) dataset covers three buildings over a three-year period, comprising more than ten thousand timeseries data points with hundreds of unique ontologies. Moreover, the metadata is standardized using the Brick schema.

This is the official repository of our NeurIPS 2024 DB Track paper that present this dataset.

Acknowledgement: This is part of the NSW, Australia Digital Infrastructure Energy Flexibility (DIEF) project.

Abstract

Buildings play a crucial role in human well-being, influencing occupant comfort, health, and safety. Additionally, they contribute significantly to global energy consumption, accounting for one-third of total energy usage, and carbon emissions. Optimizing building performance presents a vital opportunity to combat climate change and promote human flourishing. However, research in building analytics has been hampered by the lack of accessible, available, and comprehensive real-world datasets on multiple building operations. In this paper, we introduce the Building TimeSeries (BTS) dataset. Our dataset covers three buildings over a three-year period, comprising more than ten thousand timeseries data points with hundreds of unique ontologies. Moreover, the metadata is standardized using the Brick schema. To demonstrate the utility of this dataset, we performed benchmarks on two tasks: timeseries ontology classification and zero-shot forecasting. These tasks represent an essential initial step in addressing challenges related to interoperability in building analytics. Access to the dataset and the code used for benchmarking are available here: https://github.com/cruiseresearchgroup/DIEF_BTS

Dataset

View the data card here.

Snippet

For ease, we provided a very small snippet of the dataset: DIEF_B_Snippet50_3weeks.pkl.zip. Only 50 timeseries, all from Site B, haphazardly selected. This file is less than 10 MB, small enough to be sent as email attachment in most system. Accompanied with it is a short code to extract and visualize the dataset: DIEF_inspect_Snippet.ipynb.

We also provided the building metadata Site_B.ttl in the form of a Brick turtle file. Accompanied with it is a Brick definition file Brick_v1.2.1.ttl and a short code to extract the statistics: DIEF_inspect_brick.ipynb. If you have not, you will need to install the rdflib python package.

Access

Access the training dataset on FigShare.

DOI: 10.6084/m9.figshare.25912180

Note that, as of now, only the training data are made publicly available as we are planning to host a competition using this dataset. (Last update 2024 06 12)

Files description

List of files available now:

• train.zip is the raw time series data. This is a zip of a folder of pickle file. Inside each pickle file is a NumPy array with dimension [2,n] where n is the number of timesteps. The first row is the timestamp and the second row is the value.
• Site_B.ttl is the turtle file that contains the metadata of Site B using the Brick schema.
• mapper_TrainOnly.csv contains the information that maps the timeseries filenames in train.zip to the StreamID in Site_B.ttl. Some buildings in train.zip are from Site A and the brick schema is not yet made available. This is on purpose.
• train_X.zip and train_Y.csv are generated from train.zip using the function in the xySplit.py file.
• train_Y.csv see above.

The following files are missing and it is intentional:

• trainAll_2006.pkl
• trainAll_5564.pkl
• trainAll_6224.pkl
• trainAll_6486.pkl
• train_X10579.pkl
• train_X22414.pkl
• train_X26831.pkl
• train_X14584.pkl
• train_X27109.pkl
• train_X25552.pkl
• train_X30482.pkl
• train_X31456.pkl
• train_X3406.pkl
• train_X8933.pkl
• train_X14736.pkl
• train_X31557.pkl

List of files that will be made available after the competitions:

• test.zip
• Site_A.ttl
• Site_C.ttl
• mapper.csv

Code

The 20240530_class_code folder contains the code to reproduce the classification results. Here is a short description of each file:

• requirement.txt is coming, but all the information are already available in e12_pbs_py.sh.o116050675.
• xySplit.py Use this to split train.zip to train_X.zip and train_Y.csv.
• e05_naieve_02.ipynb Run this notebook to get the naive results.
• e07_02_LR.ipynb Run this notebook to get the Logistic Regression results.
• e08_02_RF.ipynb Run this notebook to get the Random Forest results.
• e09_01_XGBoost.ipynb Run this notebook to get the XGBoost results.
• thuml_tslib_dief This folder contains the modified library from Tsinghua University Machine Learning Group's library: https://github.com/thuml/Time-Series-Library
• e12_pbs_py.sh This is the setup to run the Transformer model. Consider this as the main function.
• e12_pbs_py.sh.o116050675 This is the output of the above setup. It contains detailed information about installed Python packages, their version, as well as the hardware specifications.
• e13_pbs_py.sh DLinear
• e15_pbs_py.sh PatchTST
• e17_pbs_py.sh Informer

The zeroshot_preprocess folder contains the code to preprocess the zero-shot forecasting data. Run zeroshot_preprocess/SiteA_preprocessing.ipynb to generate the processed BTS_A data.

• Note that, as of now, only the training data for the classification benchmark are made publicly available as we are planning to host a competition using this dataset. None of the data for the zero-shot forecasting benchmark are made available as of now. (Last update 2024 06 12)

The 20240612_zeroshot_code folder contains the code to reproduce the zero-shot forecasting results. This folder is a modified library from Tsinghua University Machine Learning Group's library: https://github.com/thuml/Time-Series-Library

• The zero-shot forecasting task shares the same requirement.txt as the classification task.
• ./scripts/sample_DLinear.sh is a sample script that trains a DLinear on BTS_A and tests on BTS_B and BTS_C. To run the code, replace the argument data_path with your own data path
• To train alternative models, replace the model_name with the target model name that is implemented in ./model directory and change the corresponding configurations

Brick by Brick: Automating Building Data Classification Challenge

This dataset is used as a part of the following competition: https://www.aicrowd.com/challenges/brick-by-brick-2024

A global challenge to automate building data classification, unlocking more intelligent, energy-efficient buildings for a sustainable future.

Buildings are one of the biggest energy consumers in the modern world, making energy efficiency essential. However, managing building systems data across different buildings is time-intensive and costly due to inconsistent data formats. This challenge invites you to transform building management by creating a solution that classifies building data automatically, promoting standardised, energy-efficient management for a more sustainable world.

Relevant starter kit code for the competitions are avaiable here: /competition1

Citation

Arian Prabowo, Xiachong Lin, Imran Razzak, Hao Xue, Emily W. Yap, Matthew Amos, Flora D. Salim. BTS: Building Timeseries Dataset: Empowering Large-Scale Building Analytics, 2024. arXiv:2406.08990. https://doi.org/10.48550/arXiv.2406.08990.

BibTeX:

@inproceedings{prabowo2024bts,
      title={BTS: Building Timeseries Dataset: Empowering Large-Scale Building Analytics},
      author={Arian Prabowo and Xiachong Lin and Imran Razzak and Hao Xue and Emily W. Yap and Matthew Amos and Flora D. Salim},
      year={2024},
      booktitle={Advances in Neural Information Processing Systems}
}

BibTeX:

@article{prabowo2024btsTrain,
    author = "Arian Prabowo and Xiachong Lin and Imran Razzak and Hao Xue and Emily Yap and Matthew Amos and Flora Salim",
    title = "{BTS: Building Timeseries Dataset: Empowering Large-Scale Building Analytics (TRAIN ONLY)}",
    year = "2024",
    month = "6",
    url = "https://figshare.com/articles/dataset/BTS_Building_Timeseries_Dataset_Empowering_Large-Scale_Building_Analytics_TRAIN_ONLY_/25912180",
    doi = "10.6084/m9.figshare.25912180.v1"
}

Misc

• Contact for this code repo: https://www.arianprabowo.com/
• Similar dataset: LBNL Building 59 A three-year building operational performance dataset for informing energy efficiency
• Similar dataset: Mortar (Access is currently restricted. Last update:2024 08 19)
• Other buildings related dataset: Building Data Genome Directory
• Learn more about building analytics and data-driven smart buildings from IEA EBC Annex 81.