{
  "slug": "bts-2024",
  "title": "BTS: Building Timeseries Dataset",
  "aliases": [
    "Building TimeSeries",
    "Building Timeseries Dataset",
    "DIEF BTS"
  ],
  "canonical_source_url": "https://doi.org/10.6084/m9.figshare.28705559",
  "official_documentation_url": "https://github.com/cruiseresearchgroup/DIEF_BTS",
  "introducing_paper_url": "https://proceedings.neurips.cc/paper_files/paper/2024/hash/f0430903a14db90e5ce96f101902d6d7-Abstract-Datasets_and_Benchmarks_Track.html",
  "openreview_url": "https://openreview.net/forum?id=6cCFK69vJI",
  "neurips_poster_url": "https://neurips.cc/virtual/2024/poster/97839",
  "paper_status": "accepted poster at the NeurIPS 2024 Datasets and Benchmarks Track",
  "dataset_type": "real-world irregular building-management-system telemetry with Brick 1.2.1 semantic building graphs",
  "domain": "commercial and institutional building operations in Australia",
  "release_year": 2024,
  "full_raw_release": {
    "doi": "10.6084/m9.figshare.28705559",
    "version": 3,
    "posted": "2025-04-03",
    "landing_page_download_size": "17.68 GB",
    "license": "CC BY 4.0"
  },
  "competition_release": {
    "name": "AIcrowd Brick by Brick 2024 dataset",
    "doi": "10.6084/m9.figshare.28720391",
    "version": 2,
    "posted": "2025-07-16",
    "landing_page_download_size": "11.63 GB",
    "license": "CC BY 4.0"
  },
  "repository_snapshot": {
    "url": "https://github.com/cruiseresearchgroup/DIEF_BTS",
    "commit": "ad1f0d4a6a7405b066fa9470e161cbc3b40e2833",
    "commit_authored_at": "2025-09-29T11:33:28+10:00",
    "root_license": "MIT",
    "archived": false,
    "release_tags_observed": [],
    "github_releases_observed": []
  },
  "statistics": {
    "buildings": 3,
    "timeseries": 14547,
    "timestamp_value_observations": 2863795583,
    "data_card_size": "18.77 GB",
    "start_date": "2021-01-01",
    "end_date": "2024-01-18",
    "site_timeseries_counts": {
      "BTS_A": 8349,
      "BTS_B": 851,
      "BTS_C": 5347
    },
    "site_locations": {
      "BTS_A": "Canberra, ACT, Australia",
      "BTS_B": "Newcastle, NSW, Australia",
      "BTS_C": "Melbourne, VIC, Australia"
    }
  },
  "temporal_structure": "irregular, asynchronous per-stream timestamp-value observations; benchmark pipelines optionally resample to 4-hour or 10-minute grids",
  "data_files": [
    "Site_Aaa.zip, Site_Baa.zip, Site_Caa.zip: one pickle payload per stream",
    "Site_A_metadata.csv, Site_B_metadata.csv, Site_C_metadata.csv: per-stream statistics and Brick labels keyed by StreamID",
    "Site_A.ttl, Site_B.ttl, Site_C.ttl: Brick 1.2.1 RDF/Turtle building semantic models"
  ],
  "channel_and_graph_context": [
    "Brick point class",
    "equipment",
    "location",
    "sensor",
    "setpoint",
    "command",
    "status",
    "alarm",
    "parameter",
    "RDF relationships among building entities"
  ],
  "tasks": [
    "hierarchical multi-label time-series ontology classification",
    "cross-building zero-shot forecasting",
    "domain-shift evaluation",
    "long-tail and class-imbalance evaluation",
    "graph-conditioned time-series representation learning",
    "building data interoperability"
  ],
  "published_benchmark_protocols": {
    "classification": "2-, 4-, and 8-week chunks; development and test split by time and building; hierarchical labels use positive superclasses, masked subclasses, and negative unrelated classes",
    "zero_shot_forecasting": "one month of 10-minute-resampled data; 12-step history and 12-step horizon; train on one building and test on the others"
  },
  "infrastructure": {
    "upstream_platform": "CSIRO Data Clearing House",
    "ingestion": "building-management-system telemetry uploaded using MQTTS",
    "semantic_model": "expert-authored Brick 1.2.1 graph linked to streams",
    "public_archive": "Figshare",
    "benchmark_repository": "GitHub notebooks, scripts, modified Time-Series-Library copies, sample data, data card, slides, posters, and competition artifacts",
    "competition": "Brick by Brick 2024 on AIcrowd with permanent archive and published winner solutions",
    "follow_on": "FlexTrack Challenge 2025 for demand-response flag and capacity prediction on separate digital-twin-generated data"
  },
  "action_conditioned_world_model_fit": "passive near-miss with unusually rich control-input-like semantics; Command and Setpoint streams require validated timing, target, execution status, NOOP, outcome, and intervention semantics before use as typed actions or control inputs",
  "known_limitations": [
    "only three non-residential buildings in Australia",
    "anonymization removes building layouts, occupancy patterns, schedules, and other useful context",
    "paper warns that re-identification risk cannot be ruled out when combined with external data",
    "raw data intentionally preserves realistic errors and missingness",
    "repository is an archival research codebase rather than a versioned installable package",
    "classification and forecasting benchmarks use substantial resampling and preprocessing that hide part of the raw irregular-time structure",
    "no clean reward, NOOP, action-execution, or counterfactual interface"
  ],
  "metadata_drift": [
    "paper reports 240 unique classes, repository data card reports 215 unique Brick classes, and the competition uses 94 modified Brick point subclasses",
    "paper/data-card snapshot reports 18.77 GB while the current Figshare v3 landing page reports 17.68 GB",
    "historical June 2024 README access caveats are superseded by the April 2025 full raw release"
  ],
  "license_note": "full raw and competition datasets are CC BY 4.0; repository root is MIT; competition starter material and winner-submission terms have additional artifact-specific conditions",
  "access_note": "open Figshare releases and public GitHub documentation; this knowledge base stores metadata only and does not mirror payloads",
  "created": "2026-08-10",
  "updated": "2026-08-10"
}