Research map / Large language models / AI and society
Building Timeseries Dataset: Empowering Large-Scale Building Analytics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the Building TimeSeries (BTS) dataset of real-world building operation timeseries from three Australian buildings, Brick-standardized, and benchmarks ontology classification and zero-shot forecasting.
Problem
Building analytics research lacks accessible, comprehensive, real-world multi-building operational timeseries datasets with standardized metadata, hindering interoperability and transfer across buildings.
Approach
A publicly released, anonymized, Brick-schema-labelled multivariate timeseries dataset from three buildings over three years with hundreds of point ontologies, plus baseline benchmarks for ontology multi-label classification and cross-building zero-shot forecasting.
Key results
- On timeseries ontology multi-label classification, naive baselines reach ~0.85 accuracy but near-zero F1 and mAP, and deep models (e.g., PatchTST F1 0.221, Transformer mAP 0.152) only slightly beat naive methods, indicating extreme imbalance.
- In zero-shot forecasting, PatchTST achieves the lowest SMAPE in several cross-building settings (e.g., 29.21 training on BTS_C and testing on BTS_B... specifically 28.93 trained on BTS_B tested on BTS_C at 29.21), while DLinear achieves the highest R2 in some settings (e.g., 0.545 trained on BTS_C t
- The dataset has 8349, 851, and 5347 timeseries in BTS_A, BTS_B and BTS_C respectively, versus 337 in LBNL59, and about 240 unique ontologies.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use deep learning models rather than naive methods | Timeseries ontology classification on BTS · Timeseries ontology classification | F1: 0.12955 -> 0.22162 | Helped |
| Use deep learning models rather than naive methods | Timeseries ontology classification on BTS · Timeseries ontology classification | accuracy: 0.85918 -> 0.28384 | Hurt |
| Use PatchTST rather than other forecasting baselines | Zero-shot forecasting across BTS buildings · Cross-building forecasting | SMAPE | Helped |
| Use DLinear rather than other forecasting baselines | Zero-shot forecasting across BTS buildings · Cross-building forecasting | R2 | Helped |
| Train models on BTS_A rather than BTS_B or BTS_C | Zero-shot forecasting models · Cross-building forecasting | forecasting performance | Hurt |
Surprising findings
- Naive baselines (e.g., predicting the mode or zero) achieve higher accuracy than all learned models on ontology classification, reflecting extreme class imbalance.
- Deep learning models in zero-shot forecasting trained on BTS_A transfer poorly to other buildings, attributed to its greater entity and series heterogeneity.
Limitations
- Only three buildings in Australia limits geographic and climate diversity.
- Anonymization may remove context such as layouts, occupancy and schedules, and residual re-identification risk remains.
- Benchmarks are shallow with no hyperparameter optimization.
Experimental setup
Datasets: BTS (Building TimeSeries) LBNL59
Benchmarks: BTS
Metrics: accuracy F1 mAP precision recall SMAPE R2 MAE
Baselines: Zero Random Uniform Random Proportional Mode LR XGBoost Transformer Informer DLinear PatchTST iTransformer
Builds on: Brick schema TSLib DLinear PatchTST
Closest papers by meaning
- NeuralBES: A Differentiable, Control-Aware Emulator for Scalable Building Energy Modeling (NeurIPS 2026 · similarity 0.82)
- Time Weaver: A Conditional Time Series Generation Model (ICML 2024 · similarity 0.81)
- A Closer Look at Transformers for Time Series Forecasting: Understanding Why They Work and Where They Struggle (ICML 2025 · similarity 0.81)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.80)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.80)
- This Time is Different: An Observability Perspective on Time Series Foundation Models (NeurIPS 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.