BTS: Building Timeseries Dataset

Summary

BTS is a real-world irregular building-telemetry dataset from three anonymized non-residential Australian buildings. It combines 14,547 time series and about 2.864 billion timestamp-value observations with Brick 1.2.1 semantic graphs of equipment, locations, points, and relationships.

Official Artifacts

Dataset Shape

BTS stores each stream as timestamps plus numeric values, associates it with a StreamID, and links that identifier to per-building CSV metadata and RDF/Turtle Brick graphs. The channel inventory includes sensors, setpoints, commands, statuses, alarms, parameters, equipment, and locations. Raw timestamps are irregular; published benchmark pipelines resample them.

Role In The Wiki

BTS is a strong public substrate for:

  • cross-building channel and ontology transfer;
  • graph-conditioned time-series modeling;
  • high-channel irregular telemetry;
  • long-tail and imbalance studies;
  • testing whether semantic channel context improves forecasting and representation learning.

It is a passive near-miss for action-conditioned world models. Command and Setpoint classes are control-input-like, but the release does not establish an aligned action/execution/outcome interface.

Infrastructure Assessment

The surrounding ecosystem is unusually valuable: CSIRO Data Clearing House ingestion and semantic modeling, versioned Figshare releases, inspection notebooks, benchmark code, a concluded AIcrowd challenge, a permanent competition archive, published winning solutions, and a separate FlexTrack follow-on challenge.

The GitHub repository itself is an archival research bundle rather than a maintained installable benchmark package. Reproductions should pin the exact Git commit, Figshare version, ontology, split, preprocessing, environment, and metric.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in bts-2024 rather than duplicating verdict rows here.

At the entity level, BTS is most valuable as a benchmark substrate for typed channel context, graph structure, irregular time, cross-system transfer, and long-tail telemetry. The missing step is to use those artifacts in a model interface that handles changing channel inventories and, eventually, validated control inputs and outcomes.