Source-linked AI summary
The UK-DALE dataset, domestic appliance-level electricity demand and whole-house demand from five UK homes
Jack Kelly, William Knottenbelt
TL;DR
Researchers need appliance-level ground truth alongside whole-house demand to develop energy-disaggregation algorithms. This paper presents UK-DALE, a high-resolution dataset from five UK homes, including a 655-day recording, together with a low-cost collection system.
Problem
Existing monitoring systems lacked the capacity, temporal resolution, or affordability needed to record whole-house demand and many individual appliances for disaggregation research.
Method
The authors construct UK-DALE by recording appliance and whole-house demand across five homes using a low-cost, high-resolution monitoring system.
Results
UK-DALE records appliance and whole-house demand every six seconds, with 16 kHz whole-house waveform data in three homes and 655 days of recording in House 1.
Takeaways & Limitations
UK-DALE provides open-access, high-temporal-resolution UK data containing whole-house demand and appliance-level measurements for disaggregation research.
Takeaways & Limitations
In House 1, only 80% of energy was submetered, partly because the individual appliance meters’ approximately 50 W consumption was not measured.
Abstract
from arXiv · showhide
Many countries are rolling out smart electricity meters. These measure a home's total power demand. However, research into consumer behaviour suggests that consumers are best able to improve their energy efficiency when provided with itemised, appliance-by-appliance consumption information. Energy disaggregation is a computational technique for estimating appliance-by-appliance energy consumption from a whole-house meter signal. To conduct research on disaggregation algorithms, researchers require data describing not just the aggregate demand per building but also the `ground truth' demand of individual appliances. In this context, we present UK-DALE: an open-access dataset from the UK recording Domestic Appliance-Level Electricity at a sample rate of 16 kHz for the whole-house and at 1/6 Hz for individual appliances. This is the first open access UK dataset at this temporal resolution. We recorded from five houses, one of which was recorded for 655 days, the longest duration we are aware of for any energy dataset at this sample rate. We also describe the low-cost, open-source, wireless system we built for collecting our dataset.
1. Background & Summary
Residential energy users often misestimate whole-house and appliance consumption, while appliance-level feedback could support more efficient behaviour. The paper presents UK-DALE, an open-access UK dataset capturing appliance and whole-house demand at high temporal resolution across five houses.
- Motivation: Residents often misestimate whole-house and individual-device energy consumption, underestimating heating and overestimating salient devices such as lights and televisions.These inaccurate estimates may lead to higher energy consumption.
- Motivation: Energy use can differ by two or three times among otherwise similar houses, with differences attributed to occupant consumption behaviour.Better feedback about device-level consumption could help users adjust appliance use more efficiently.
- Problem: Smart meters measure whole-house consumption, whereas behavioural research suggests consumers manage electricity best with appliance-by-appliance information.Energy disaggregation estimates appliance-level consumption from a whole-house smart-meter signal.
- Research need: Researchers need large, open-access field datasets to develop disaggregation algorithms without each researcher recording a separate dataset.Country-specific data matters because appliance sets and usage patterns vary significantly between countries.
- Contribution: UK-DALE records five houses, measuring individual-appliance active power and whole-house apparent power every six seconds, with House 1 recorded for 655 days.In three houses, whole-house voltage and current were sampled at 44.1 kHz and down-sampled to 16 kHz for storage; active power, apparent power, and RMS voltage were calculated at 1 Hz.
2. Methods
The methods combine appliance-level monitoring with whole-house metering designed to approximate smart-meter data. The authors built custom low-cost hardware to support dense, high-frequency collection while addressing measurement and installation constraints.
- Dataset requirements: The dataset design required simultaneous appliance-level and whole-house active-power recording, sampling at least every 10 seconds, and long-duration collection.These attributes support validating or training disaggregation systems and using whole-house demand as algorithm input.
- Appliance monitoring: Because UK mains rings serve multiple sockets, individual appliances were measured using plug-in individual appliance monitors installed between each appliance and its wall socket.The approach monitored individual appliances once every 6 seconds.
- Appliance monitoring: The commercial EcoManager base station supported only 14 plugs and one-minute data, so the researchers built a custom base station for up to 54 appliances with 10-second-or-faster sampling.The custom station used a Nanode platform with an ATmega328P microcontroller and RFM12b radio module.
- Appliance monitoring: Hard-wired appliances were monitored with Current Cost transmitters and current-transformer clamps, which transmitted packets every 6±0.3 seconds without collision avoidance or retransmission.The base station reduced collision risk by avoiding transmission shortly before expected transmitter packets.
- Whole-house metering: The whole-house system recorded active power once per second and voltage and current waveforms at 44.1 kHz using a computer sound card, CT clamp, and AC-AC adapter.The system saved active power, apparent power, and RMS voltage once per second, with relative errors consistently below 2%.
- Measurement limitations: CT-clamp power estimates can vary by +20% to -12% with UK mains-voltage variation, creating problematic demand changes for disaggregation algorithms.The transmitters generally hard-code voltage rather than measuring it, while constant-power devices are not affected in the same way.
3. Data Records
Section 3 describes UK-DALE’s directory structure, appliance and wiring metadata, and three available data resolutions. The 16 kHz whole-house waveform recordings use compressed, hour-sized FLAC files and require substantial storage.
- Data organization: UK-DALE contains five house directories, each with channel CSV files for electricity meters and a labels.dat mapping from channel numbers to appliance names.All CSV files use a single space as the column separator, following REDD.
- Metadata: Detailed YAML metadata follows the NILM Metadata schema and describes appliance specifications, meter wiring, measurements, and appliance rooms.This metadata is a stated difference from REDD.
- Available data: UK-DALE provides 6 second, 1 second, and 16 kHz data, available through the UK Energy Research Council’s Energy Data Centre and a project website.The website is intended to be updated as more data are collected.
- 16 kHz data: The complete 16 kHz dataset requires 4 TBytes of storage and is supplied as 200 MByte files, each recording 1 hour of data.The 16 kHz data are recorded by the sound card power meter.
- 16 kHz data: For houses 1, 2, and 5, UK-DALE stores whole-house current and voltage waveforms as stereo 16 kHz FLAC files split into hour-sized chunks.File names use the form vi-<T>.flac, where T is a UNIX timestamp with micro-second precision marking the recording start.
4. Technical Validation
Technical validation shows that UK-DALE captures appliance and whole-house demand with interpretable temporal and power-demand patterns. Independent comparisons indicate low measurement error for the sound card power meter, while Current Cost CT accuracy depends on load level.
- Dataset validation: House 1’s typical-day plot separates whole-house mains demand from the top-five appliances and other submeters, with the remaining gap indicating unsubmetered energy.The top five appliances are ranked by energy consumption, and the residual between mains demand and submetered demand is small.
- Dataset validation: Appliance power histograms expose device-specific operating states, including fridge peaks near 90 W, 17 W, and 250 W and six discrete vacuum-cleaner settings.The fridge peaks correspond respectively to compressor operation, the lamp, and defrosting.
- Meter validation: Less than 2% relative error was measured for the sound card power meter, while the Current Cost CT meter stayed below 6% above 100 watts.The whole-house power demand very rarely drops below 100 watts, the condition under which the CT result applies.
- Meter validation: 2.71% was the relative difference between House 1 sound-card energy and the utility meter, which recorded 4030.60 kWh versus 4142.93 kWh.The comparison used a period with continuous sound-card recording and utility-meter readings.
- Meter validation: 1.79% in House 1, 7.30% in House 2, and 5.68% in House 5 were the relative differences between sound-card and Current Cost whole-house meters.The comparison covered apparent energy recorded by the sound card power meter against the Current Cost meter in houses with two mains meters.
5. Usage Notes
UK-DALE can be processed by REDD-compatible software and includes NILMTK support through an importer and downloadable HDF5 version. Users should preprocess for packet loss, timing drift, monitor power gaps, and appliance-specific on/off thresholds.
- Software support: REDD-compatible software should be able to open UK-DALE files, although it will ignore UK-DALE’s extra metadata.
- Software support: NILMTK includes a UK-DALE importer, handles its metadata, and provides a downloadable HDF5 dataset version.
- Pre-processing considerations: Around 6% of Current Transformer packets and 0.02% of Individual Appliance Monitor packets are lost, while 6-second sample periods may drift by one second.
- Pre-processing considerations: Gaps longer than two minutes generally indicate that an appliance and its monitor were switched off from the mains, including when appliances are physically unplugged.
- Pre-processing considerations: A 5-watt threshold typically distinguishes appliance on/off states, but metadata records an alternative on power threshold when an appliance differs.
7. Author Contributions
Jack Kelly led UK-DALE’s technical development and wrote most of the paper, while William Knottenbelt provided supervision, recruitment support, and editorial feedback.
- Jack Kelly built the hardware, wrote the software, and wrote the majority of the paper.
- William Knottenbelt supervised Jack Kelly’s PhD, helped recruit MSc students, and provided conceptual and editorial guidance.
9. Data Citations
The section lists two citations by Kelly and Knottenbelt, each associated with a distinct UKERC Energy Data Centre DOI from 2015.
- 1. Kelly and Knottenbelt are cited with UKERC Energy Data Centre DOI:10.5286/UKERC.EDC.000001 (2015).
- 2. Kelly and Knottenbelt are cited with UKERC Energy Data Centre DOI:10.5286/UKERC.EDC.000002 (2015).