---

tags: - energy disaggregation - non-intrusive load monitoring - time series - electrical load monitoring license: unknown # License information was not explicitly stated in the paper, might need clarification. language: - en pretty_name: BLUED (Building-Level fUlly-labeled dataset for Electricity Disaggregation)

Dataset Card for BLUED

Dataset Description

BLUED (Building-Level fUlly-labeled dataset for Electricity Disaggregation) is a public dataset designed for event-based Non-Intrusive Load Monitoring (NILM) research. It contains high-frequency voltage and current measurements from a single-family home in the United States over one week. The key feature of this dataset is the detailed labeling of appliance state transitions (events), providing ground truth for evaluating event-based disaggregation algorithms. The dataset aims to facilitate the development, testing, and comparison of NILM algorithms.

Dataset Details

• Data Collection:
  • Data was collected over one week in October 2011 from a single-family house in Pittsburgh, Pennsylvania.
  • Aggregate voltage and current measurements were captured at the main distribution panel using a National Instruments DAQ (NI USB-9215A) at a sampling rate of 12 kHz. Current was measured using split-core current transformers, and voltage was measured using a voltage transformer.
  • Ground truth for appliance events was collected using a combination of plug-level power meters (FireFly sensors), environmental sensors (light, sound, vibration, etc.), and circuit-level current measurements.
  • Events were defined as changes in power consumption greater than 30 watts lasting at least 5 seconds.
  • Timestamps for ground truth events were manually synchronized with the aggregate power signal via visual inspection.
• Data Content:
  • Raw voltage (one phase) and current (two phases) waveforms sampled at 12 kHz.
  • Computed active power at 60 Hz.
  • A list of timestamped events, identifying the appliance and the transition type (e.g., on/off).
  • Covers approximately 50 electrical appliances, though not all were active or met the event criteria during the collection week.
  • Includes 2,482 labeled events in total, with 2,355 attributed to known appliances and 127 from unknown sources (clustered into potentially 11 distinct appliances). Events are split between Phase A (904 events) and Phase B (1578 events).
• Data Format: Raw current and voltage files, along with a list of event timestamps. Active power computed at 60Hz is also included.
• Data Splits: The paper presents preliminary results using the whole week but suggests future work might involve splitting into training/testing sets.

Uses

• Non-Intrusive Load Monitoring (NILM): Primarily designed for developing and evaluating event-based energy disaggregation algorithms.
• Appliance Usage Pattern Analysis: Studying how and when different appliances are used in a residential setting.
• Occupancy Detection: Inferring household occupancy based on appliance usage.
• Energy Management & Efficiency: Developing strategies for residential energy savings.
• Anomaly Detection & Fault Diagnostics: Identifying unusual appliance behavior or potential faults.
• Assisted Living Applications: Monitoring activities of daily living through appliance usage.

Dataset Limitations

• Duration: One week of data may not capture the usage patterns of all appliances, especially seasonal ones (like the air conditioner) or those used infrequently (like the dryer).
• Sensor Frequency Limitation: The current sensors used had a cutoff frequency around 300 Hz, limiting the analysis of higher-frequency harmonics (beyond the 5th harmonic).
• Incomplete Ground Truth: Approximately 5% of events detected in the aggregate signal could not be attributed to the monitored appliances and are labeled as "unknown". Some appliances (~25%) had no registered events meeting the criteria during the collection week.
• Single Home: Data represents only one specific home and its occupants' behavior.

Citation

@inproceedings{anderson2012blued,
  title={BLUED: A fully labeled public dataset for event-based non-intrusive load monitoring research},
  author={Anderson, Kyle and Ocneanu, Adrian and Benitez, Diego and Carlson, Derrick and Rowe, Anthony and Berg{\'e}s, Mario},
  booktitle={Proceedings of the 2nd ACM SIGKDD international workshop on data mining applications in sustainability},
  pages={1--8},
  year={2012},
  organization={ACM}
}