# BTS: Building Timeseries Dataset Canonical dataset DOI: Official repository and documentation: Introducing paper: OpenReview: Wiki source: [BTS](../../wiki/sources/bts-2024.md) ## Dataset Type BTS is a real-world building-management-system telemetry dataset from three anonymized non-residential buildings in Australia. It combines irregular numeric time series with a Brick 1.2.1 semantic model of building locations, equipment, points, and their relationships. ## Release And Version Snapshot - Paper release: NeurIPS 2024 Datasets and Benchmarks Track, poster. - Full raw data: Figshare DOI `10.6084/m9.figshare.28705559`, version 3, posted 2025-04-03; the current landing page reports a 17.68 GB download. - Competition data: Figshare DOI `10.6084/m9.figshare.28720391`, version 2, posted 2025-07-16; the current landing page reports an 11.63 GB download. - Repository snapshot audited: commit `ad1f0d4a6a7405b066fa9470e161cbc3b40e2833`, authored 2025-09-29. - This knowledge-base ingest records metadata only and does not mirror the raw data, snippets, competition archives, notebooks, PDFs, or winning submissions. ## Dataset Shape The repository data card reports: - 3 buildings; - 14,547 time series; - 2,863,795,583 timestamp-value observations; - 18.77 GB in its older dataset-size snapshot; - coverage from 2021-01-01 through 2024-01-18, with site-specific end dates; - real telemetry that was intentionally not cleaned so realistic errors, missingness, and unusual values remain available for study. The paper's site-level table reports 8,349, 851, and 5,347 time series for BTS A, B, and C respectively. Individual streams can have different timestamps and sampling rates. ## Data And Metadata Structure The full-release README describes three artifact families: - `Site_*aa.zip`: one pickle file per stream, containing a stream identifier, a one-dimensional timestamp array, and a one-dimensional value array; - `Site_*_metadata.csv`: per-stream basic statistics and Brick class, keyed by `StreamID`; - `Site_*.ttl`: Turtle/RDF building semantic models that link time-series points to equipment, locations, and other building entities through Brick 1.2.1. The repository also includes a 50-stream, three-week BTS B snippet and inspection notebooks. Pickle payloads SHOULD be loaded only from the official versioned release or another trusted source because Python pickle deserialization is executable. ## Ontology And Channel Roles The Brick metadata includes sensors, setpoints, commands, status points, alarms, parameters, equipment, and locations. These labels provide unusually rich channel context and graph structure for real building telemetry. They do not automatically provide a clean action channel. A `Command` or `Setpoint` point is control-input-like, but a world-model dataset still needs aligned action timing, target, parameters, execution status, `NOOP` semantics, and downstream outcomes before those streams can be treated as typed actions or interventions. ## Tasks And Benchmarks ### Multi-label ontology classification The paper segments streams into 2-, 4-, or 8-week chunks and predicts a hierarchical Brick label vector. A true point class and its superclasses are positive, subclasses are masked, and unrelated labels are negative. Development and test data are split by both time and building. The repository contains naive, classical-ML, and modified Time-Series-Library baselines. ### Cross-building zero-shot forecasting The paper also reports an appendix benchmark that resamples irregular streams to 10-minute intervals, uses a 2-hour history and 2-hour horizon, trains on one building, and tests on the others. DLinear is evaluated as a multivariate model; PatchTST, Informer, and iTransformer are handled as univariate models because their tested input interfaces do not support building-to-building channel-count changes. ### Brick by Brick challenge The concluded AIcrowd competition turns BTS into a hierarchical multi-label time-series classification challenge with 94 point subclasses, hidden test partitions, an evaluation script, a starter kit, a permanent Figshare archive, a leaderboard, and published winning solutions in The Web Conference 2025 companion proceedings. ## Infrastructure Around The Dataset The dataset sits inside the NSW Digital Infrastructure for Energy Flexibility stack: 1. Building-management systems produce telemetry from sensors, meters, devices, and control-related points. 2. CSIRO's Data Clearing House ingests the streams over MQTTS and links them to a semantic building model. 3. Expert engineers map building points and relationships into Brick 1.2.1. 4. Figshare provides versioned public raw and competition releases. 5. The GitHub repository provides a data card, sample data, inspection notebooks, benchmark code, modified Time-Series-Library copies, challenge archives, and winner references. 6. The separate FlexTrack 2025 challenge extends the project ecosystem toward demand-response event detection and capacity estimation, but uses digital-twin-generated 15-minute data rather than the BTS raw telemetry. ## Repository Engineering Snapshot The repository is useful as a research archive but is not a maintained installable data package: - benchmark code is organized as dated notebook/script directories; - large copies of modified Time-Series-Library code are checked in; - the classification environment pins older packages such as PyTorch 1.7.1, NumPy 1.23.5, and pandas 1.5.3; - generated outputs, compiled Python bytecode, slides, posters, competition submissions, and papers are committed; - the audited tree has no release tags, GitHub Releases, package metadata, CI workflow, test suite, or machine-readable dataset checksum manifest. Reproduction SHOULD therefore pin the Git commit, Figshare version, exact benchmark split, preprocessing code, label ontology, dependency environment, and evaluation metric rather than treating the repository's `main` branch as a stable package release. ## Access And License Notes - The full raw dataset and archived competition dataset are published on Figshare under CC BY 4.0. - The repository root code/documentation license is MIT. - The competition folder contains additional UNSW-owned starter-kit and challenge-description material, while the challenge rules separately required prize-eligible submissions to be provided under Apache 2.0. These terms SHOULD NOT be collapsed into one repository-wide license statement. - The authors report that direct identifiers were replaced with UUIDs and exact buildings were anonymized. The paper still warns that linkage with external information can create residual re-identification risk. ## Known Limitations And Drift - Only three non-residential Australian buildings are included, so geographic, climate, regulation, building-type, and operating-practice diversity is limited. - Anonymization removes useful layout, occupancy, and schedule context. - The paper reports 240 unique classes, the repository data card reports 215 unique Brick classes, and the competition uses 94 modified Brick point subclasses. These are different artifact/protocol snapshots and MUST be pinned rather than silently reconciled. - The paper/data-card snapshot reports 18.77 GB, while the current Figshare v3 landing page reports a 17.68 GB download. This likely reflects artifact/version or size-accounting drift, but the available metadata does not prove one reconciliation. - Historical README text saying zero-shot data were unavailable in June 2024 is superseded by the full Figshare release in April 2025. - Classification winner scores are not directly comparable with the paper table because the competition uses a changed partition, 94-class label surface, and a separate evaluation protocol.