Kaggle — Appliances Energy Prediction (UCI mirror + notebooks)

The community-packaged edition of the most-cited public benchmark for household appliance-load regression: one low-energy house near Chievres Airport, Belgium, instrumented for roughly 4.5 months from 11 January 2016, producing 19,734 ten-minute rows in a single file, KAG_energydata_complete.csv.

Each row pairs the number every model is asked to predict - appliance draw in watt-hours, logged with m-bus energy meters - against everything plausibly capable of explaining it: separately metered lighting, temperature and humidity from a ZigBee network at nine positions (nodes reporting about every 3.3 minutes, averaged onto the ten-minute grid), and four weather variables merged from the airport station by timestamp. Two columns are random numbers by design - rv1 and rv2 exist purely to catch models claiming skill they do not have.

What the hosting layer adds over the raw repository drop is ecosystem rather than content: a usability rating, version tracking, 43 community notebooks and live discussion threads attached to the same bytes. The table itself belongs to the household appliances data hub alongside every other metering source on the shelf.

One row per ten-minute interval, column names arriving verbatim:

# One row per 10-minute interval; column names arrive verbatim, casing and all

date                 Appliances lights T1     RH_1   T2     RH_2   T_out RH_out Windspeed Visibility Tdewpoint rv1    rv2
2016-01-11 17:00:00  60         30     19.89  47.597 19.20  44.79  6.60  92.0   7.0       63.0       5.3       13.28 13.28
2016-01-11 17:10:00  60         30     19.89  46.693 19.20  44.72  6.48  92.0   6.67      59.17      5.2       18.61 18.61

# numerics ship as quoted strings with leading spaces - cast before any arithmetic
# rv1 and rv2 print identical values (13.28, then 18.61): one random variable carried twice,
# so pipelines treating them as independent evidence fail their own leakage audit.

Three conventions decide whether a pipeline survives contact with the full table. Numerics are quoted strings with leading whitespace, so explicit casting beats type inference. rv1 and rv2 print identical values - one random variable carried twice - so pipelines treating them as independent evidence fail their own audit. Room numbering is positional: position six sits outside on the building's north face despite sitting mid-sequence, and that outdoor pair bridges the indoor network to the weather block. Rows scoped to your named intervals ship with the sample.

Ten dictionary entries cover all 29 columns, verified against live rows during the August 2026 research pass rather than inferred from column names. Identity (date) keys every row; the target (Appliances) carries whole-house draw in watt-hours; lights is metered separately; the nine T*/RH_* pairs supply nine-position climate; T_out, RH_out, Press_mm_hg, Windspeed, Visibility and Tdewpoint merge the airport station by timestamp; and rv1/rv2 ship as deliberate noise.

The naming pattern compresses cleanly because it is systematic - T for temperature in Celsius, RH for relative humidity in percent - but read the numbering as sensor position, not alphabetization. Every definition above holds on the exact release the host serves today; nothing in the table has moved since September 2017.

Three chips summarize the footprint:

• Geography - a single low-energy house near Chievres Airport, Belgium: one building, nine instrumented positions, no multi-site panel.
• Time - 19,734 consecutive ten-minute intervals beginning 11 January 2016 and running roughly 4.5 months through late May 2016; one static release dated 16 September 2017, with no refresh schedule behind it.
• Granularity - whole-house aggregate draw per interval, nine-room climate averaged onto the same grid, airport weather merged at matching resolution.

One boundary worth knowing upfront: this is one house observed deeply, not a market observed broadly. Device-level channels and many-home panels live in the neighbouring metering sets on the same shelf - pair them when the analysis needs breadth this table cannot provide.

API, files, or your warehouse. Daily, weekly, or hourly.

The table is static, so a delivery resolves to the same rows every time - what varies is packaging: pushed to storage, served over an endpoint, or synced straight into Snowflake, BigQuery or Redshift. Casting rules for the quoted numerics travel with the extract so the schema validated in the sample is the schema the recurring feed keeps.

Name the windows, rooms or derivations you want precomputed - calendar features, hourly aggregates, long-format reshaping - and the sample arrives already scoped with the field dictionary attached and the join keys called out.

Four workloads lean on this table hardest:

1. Data scientists fork a notebook, reproduce the canonical result, then iterate - the fastest route to a defensible baseline; see data scientists use cases.
2. Educators and students teach time-series regression on well-trodden ground where the discussion threads double as office hours; see the ML model training use case.
3. Builders wire one small CSV into energy-dashboard demos and prototypes, crediting the origin study; see builders use cases.
4. Product teams reference typical load shapes when scoping smart-home features; see competitive intel product teams use cases.

One boundary: nobody ships production forecasts off one Belgian house. Everybody ships models they first proved here - which is exactly the job this table does.

Provenance note - the host is Kaggle user loveall's packaging of the UCI Appliances Energy Prediction set. Cite the originating record and Candanedo et al.'s 2017 Energy and Buildings study (DOI 10.24432/C5VC8G) rather than the mirror page when publication-grade attribution matters.

Freshness note - figures quoted here were measured at research time: roughly 12,400 pulls of the file, 204,000 page views, 157 votes and 43 linked notebooks on the host, with the kernel count reported by the metadata route conflicting with higher figures quoted elsewhere. Treat any count in a delivered extract as measured.

Dictionary note - all 29 columns are typed with worked examples drawn from the first rows, and two genuine sample rows ship with the record.

Where to go next - the raw repository record when provenance outranks convenience, REFIT Electrical Load Measurements for many homes instead of one, Pecan Street Dataport for circuit-level scale, and the rest of the shelf in best household-appliances datasets.