Household Appliances · UCI Machine Learning Repository (record ID 374)
UCI Appliances Energy Prediction Dataset
Datadory delivers uci appliances energy prediction dataset data covering 19,735 ten-minute observations of whole-house appliance energy use in watt-hours, joined with nine-room indoor temperature and humidity, four airport weather variables and two deliberate noise columns, recorded over roughly 4.5 months in a Belgian low-energy house starting 11 January 2016.
API, files, or your warehouse. Daily, weekly, or hourly.
- Where it covers
- A single low-energy house near Chievres Airport, Belgium - one building, nine instrumented positions, no multi-site panel.
- How far back
- 19,735 consecutive intervals beginning 11 January 2016 and running roughly 4.5 months through late May 2016; one canonical static release dated 14 February 2017 rather than rolling revisions.
- How fine
- Whole-house aggregate appliance energy per 10-minute interval, nine-room indoor climate channels averaged onto the same grid, airport weather merged at matching resolution.
How is the UCI Appliances Energy Prediction Dataset delivered?
API, files, or your warehouse. Daily, weekly, or hourly.
The full 29-column history ships however your stack wants it - pushed to storage, served over an endpoint, or synced straight into your database. Name the fields and the cadence when you request the sample; the sample arrives first either way, with the field dictionary attached and the join keys called out. Whatever cadence you settle on, the schema validated in the sample is the schema the recurring feed keeps.
What should I know before requesting a sample?
Benchmark scope - because every published model reports on these exact 19,735 rows, results are comparable across papers but not automatically transferable to multi-home settings; treat headline accuracies as method validation, not deployment estimates. If your question needs many homes instead of one deeply measured home, REFIT covers 20 UK households and Pecan Street spans a 2,072-home network.
Parsing discipline - numeric columns arrive as quoted strings with leading whitespace and full float precision. Document your coercion rules once; silent type inference is how the twin random variables sneak into feature importances. There are no missing values to impute, and the full table holds comfortably in memory at roughly 12 MB across 19,735 rows and 29 columns.
Citation practice - the canonical study is Candanedo, Feldheim and Deramaix, 'Data driven prediction models of energy use of appliances in a low-energy house', Energy and Buildings, Volume 140, April 2017, pages 81-97, under DOI 10.24432/C5VC8G.
Field dictionary - UCI Appliances Energy Prediction Dataset (29 columns per row)
| Field | Type | Definition | Example |
|---|---|---|---|
| date | datetime | Observation timestamp at 10-minute resolution. | "2016-01-11 17:00:00" |
| Appliances | number | Whole-house appliance energy use in watt-hours for the interval; the regression target. | 60 |
| lights | number | Light fixture energy use in watt-hours for the interval. | 30 |
| T1 / RH_1 | number | Kitchen temperature (Celsius) / humidity (percent). | 19.89 / 47.60 |
| T2 / RH_2 | number | Living room temperature / humidity. | 19.20 / 44.79 |
| T3 / RH_3 | number | Laundry room temperature / humidity. | 19.79 / 44.73 |
| T4 / RH_4 | number | Office room temperature / humidity. | 19.0 / 45.57 |
| T5 / RH_5 | number | Bathroom temperature / humidity. | 17.17 / 55.20 |
| T6 / RH_6 | number | Temperature / humidity outside the building, north side. | 7.03 / 84.26 |
| T7 / RH_7 | number | Ironing room temperature / humidity. | 17.20 / 41.63 |
| T8 / RH_8 | number | Teenager room temperature / humidity. | 18.20 / 48.90 |
| T9 / RH_9 | number | Parents' room temperature / humidity. | 17.03 / 45.53 |
| T_out | number | Outside temperature from the Chievres Airport weather station, Celsius. | 6.60 |
| Press_mm_hg | number | Outside barometric pressure in mm Hg. | 733.5 |
| RH_out | number | Outside humidity in percent. | 92.0 |
| Windspeed | number | Wind speed in m/s. | 7.0 |
| Visibility | number | Visibility in km. | 63.0 |
| Tdewpoint | number | Dew point temperature in Celsius. | 5.3 |
What teams do with it
- Appliance-load forecasting baselines Baseline any new appliance-load or short-term-load forecasting model against the result the entire literature reports on these exact rows, so claims stay comparable instead of anecdotal.
- Feature-selection and leakage studies Study feature selection honestly: the built-in random columns make leakage and false skill detectable, which is why methods papers keep returning to this record.
- Anomaly detection and disaggregation prototypes Prototype anomaly detection and disaggregation logic against nine-room climate context before committing hardware budgets to field pilots.
- Building-performance calibration Demonstrate how strongly indoor-outdoor gradients explain appliance draw, and calibrate expectations for what ambient sensing alone can and cannot recover.
- Pipeline hardening Exercise pipelines on a table small enough to reproduce end-to-end yet messy enough (quoted numerics, twin columns, positional room codes) to catch real parsing bugs.
Questions buyers ask
What does the UCI Appliances Energy Prediction Dataset include?
Twenty-nine columns per 10-minute interval: whole-house appliance energy use in watt-hours, separately metered lighting energy, temperature and humidity for nine positions in and around the house, four weather variables from the Chievres Airport station, and two deliberate random-variable columns.
How many rows and what time range does it cover?
19,735 rows at ten-minute resolution spanning roughly 4.5 months from 11 January 2016 through late May 2016, captured in a single low-energy house near Chievres Airport, Belgium.
Which rooms carry temperature and humidity sensors?
Nine positions: kitchen, living room, laundry room, office, bathroom, ironing room, teenager room, parents' room, plus one pair mounted outside on the north side of the building. Numbering follows sensor position, not alphabetical order.
Why are rv1 and rv2 in an energy dataset?
They are non-dimensional random variables inserted on purpose as a trap: any model that reports feature importance for them is leaking noise into its predictions. They double as a free audit of your pipeline's discipline.
Is the data quality good enough for modeling without cleaning?
Largely yes: no missing values, a continuous 10-minute grid and verified sensor placement. Two caveats remain - numeric values arrive as quoted strings with leading whitespace, and the twin random columns must be excluded from features rather than silently absorbed.
How large is the dataset in practice?
About 12 MB uncompressed as a single table of 19,735 rows by 29 columns - small enough to load fully into memory and reproduce end-to-end, which is part of why it became the shared benchmark it is.
Datasets that pair with this one
- REFIT Electrical Load Measurements (Cleaned) Many homes instead of one: cleaned household load recordings for 20 UK houses, the natural out-of-sample companion.
- Pecan Street Dataport - High-Resolution Household Energy & Water Circuit-level electricity at one-second grain across a 2,072-home research network - the scale play to this benchmark's depth.
- Tracebase - Appliance-Level Power Traces Outlet-level power traces organised by appliance type - device signatures this aggregate table can only imply.
- Kaggle - Appliances Energy Prediction (UCI mirror + notebooks) The same table packaged with community notebooks, useful for worked examples beside the raw rows.
- household-appliances data hub Every household appliances dataset in the Datadory catalog, compared side by side.
See the rows before you pay anything.
Name this dataset and we send real records from it — scoped to the fields you asked for.