EIA Consumption & Efficiency - Electricity Surveys (RECS/CBECS/MECS)

The demand side of the American grid, measured unit by unit instead of modeled. The U.S. Energy Information Administration runs three national consumption surveys under one roof, and this record bundles them: RECS, the Residential Energy Consumption Survey, whose 2020 wave observed 18,496 households standing in for the country's 123.5 million occupied primary residences; CBECS, the Commercial Buildings Energy Consumption Survey, which most recently reported on an estimated 5.9 million commercial buildings that consumed 6.8 quadrillion Btu and spent $141 billion in 2018; and MECS, the Manufacturing Energy Consumption Survey, whose 2022 iteration arrived on a staged release schedule running to hand-suppressed tables in March 2026.

Together they answer the questions billing data cannot: not just how much electricity a home used but what it used it for - space heating, water heating, cooling, appliances - and what kind of home was doing the using. Within Datadory's electric utilities shelf this is the only demand-side record scored at unit grain, and it scores 9/10 on our rubric against a catalog average of 7.81 across the 1,744 datasets we cataloged. Get a sample of this dataset scoped to your sectors before you build on it.

One record per surveyed household. A real row from the research capture, exactly as it arrives:

# one row per surveyed household - verbatim RECS 2020 variables

# RECS 2020 microdata, household record 1 of 18,496:
DOEID        : 100001       REGIONC      : WEST
DIVISION     : Mountain South   state_postal : NM
TYPEHUQ      : 2            BEDROOMS     : 4
TOTROOMS     : 8            KWH          : 12521.48
DOLLAREL     : 1955.06

# survey frame around the rows:
RECS 2020    : ~18,500 sampled households representing
               123.5 million occupied primary residences
CBECS 2018   : 5.9 million commercial buildings consuming
               6.8 quadrillion Btu, spending $141 billion
MECS 2022    : manufacturing establishment tables released
               on a staged schedule through March 2026

# ... repeats across 18,496 households x 799 columns, plus
#     building-level CBECS records and NAICS-by-region MECS tables

Three things worth reading off it. First, consumption rides beside cost - KWH next to DOLLAREL - so implied unit prices fall out of a division rather than a procurement exercise. Second, geography arrives at the granularity disclosure processing permits: census region and division on the row, state postal codes only where they survive masking, fuller state detail in dedicated estimate tables. Third, the survey frames beneath the row are what make it useful - 18,496 households carrying weights that expand them to 123.5 million homes, 5.9 million commercial buildings behind the CBECS figures, manufacturing establishments organized by NAICS industry in the MECS tables. Ignore the frames and you have described a sample, not the country.

Eleven field families form the household spine, defined below with examples taken straight off a real RECS 2020 row.

Identity and geography - DOEID anonymizes the respondent while keying every RECS file together; REGIONC, DIVISION and state_postal place the home at census region, division and, where allowed, state.

Structure and occupancy - TYPEHUQ codes the unit from mobile home through single-family detached to apartments by building size; BEDROOMS and TOTROOMS size it room by room; TOTSQFT and TOTUSQFT measure total and used floor area; KOWNRENT separates owned from rented.

Consumption and expenditure - KWH carries total reference-year electricity including solar self-generation, DOLLAREL carries the bill that came with it.

Weather normalization - HDD65 and CDD65 bring heating and cooling degree days to the household location, the pair that keeps a Minnesota-and-Arizona comparison honest.

Beyond the spine, the full 799-variable layout extends into equipment inventories, appliance holdings, end-use consumption and expenditure splits, and the analysis-weight machinery - all available under additional fields on request: named at sampling, delivered as typed columns.

The eleven families above anchor the core. Around them, the survey estate maps onto specific jobs at sampling:

• End-use splits - space heating, water heating, cooling, refrigeration and lighting consumption and expenditures, each resolved into typed columns beside the household row.
• Equipment and appliance inventories - heating, cooling and water-heating systems plus major-appliance holdings, keyed to the same DOEID so joins stay clean.
• State-level estimate sheets - RECS consumption and expenditure detail for the select states that publish it, split out standalone.
• Cross-cycle stitched frames - RECS 2009, 2015 and 2020 joined on consistent definitions for trend work the single vintages cannot support alone.
• CBECS and MECS series - building-characteristics and consumption-expenditure extracts, and the manufacturing first-use, end-use, fuel-switching and energy-management table series, shaped for warehouse loads.

Name the survey, sector and field families when you request the sample and it comes back shaped to that specification.

• Geography: United States national estimates with deliberately bounded detail. RECS publishes state-level consumption and expenditure estimates for select states and reports at census division otherwise; CBECS and MECS report national and regional tables. The masking is the price of publishing unit-level records at all.
• Temporal: fixed vintages rather than a ticker. RECS has been fielded since 1978 with published cycles in 1993, 1997, 2001, 2005, 2009, 2015 and 2020, and a 2024-wave initial release underway; CBECS cycles run 1992 through 2018; MECS runs 1994 through 2022. Trend work stitches snapshots on consistent definitions instead of reading a continuous series.
• Granularity: household-level microdata for RECS, building-level records for CBECS, establishment-level tables for MECS - each paired with aggregated estimate tables and the weights that scale samples to national totals.

Set it against the rest of the pool in our best electric utilities datasets ranking, where it leads as the only demand-side record: Form EIA-860 inventories the generators, Electric Power Annual and Monthly aggregate the flows, and SEDS fills the state map - none of them describe the consuming unit.

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

Survey-shaped records reward whole-vintage handling. Take complete RECS or CBECS extracts sized for overnight warehouse loads, cut slices by census division, housing type or end use when one peer group is all that matters, or run a feed keyed on survey plus vintage so a new cycle lands as fresh rows beside the old ones instead of overwriting them.

Cadence follows the question. Hourly suits dashboards waiting on a fresh vintage, nightly loads suit enrichment pipelines, weekly pulls suit research snapshots. Every delivery ships the complete field dictionary above, value-label references for the coded columns, and validation rows so your tabulations reconcile against the published national estimates before anyone downstream asks.

• Residential load modeling - appliance holdings, dwelling traits and billed kilowatthours arrive on one row, ready for feature work and synthetic load curves. See data science use cases.
• Efficiency and electrification market sizing - counts of homes by structure type, tenure and region, plus the commercial building stock beside them, put denominators under every market model. (market sizing)
• Building benchmarking - CBECS intensity distributions by principal activity, size band and division give ESG and portfolio work an honest peer group. (ESG emissions analysis)
• Sector demand forecasting - MECS end-use tables ground industrial scenarios in what each manufacturing industry actually consumed; RECS does the same for residential classes. (demand forecasting)
• Citation-grade research - documented questionnaires and stated sample sizes make every figure quotable. (citation-grade research)

Data scientists and ML engineers get the strongest panel in the catalog: weighted, weather-normalized households with hundreds of labeled variables and documentation honest about how each estimate was built.

Market researchers and consultants segment with official denominators already counted - households, buildings, floorspace, establishments - rather than commissioning a survey to relearn them.

Investors and quants read electrification and efficiency exposure out of the same tables regulators publish, before sizing positions in utilities and equipment makers.

Developers and builders build sustainability dashboards on a codebook-documented column vocabulary that survives between survey cycles. Journalists, academics and students quote the measurement itself, with methodology documented well enough to footnote.

It is the outlier on the shelf, and the shelf needs it. Every other record in the electric utilities pool looks at the system from the supply side: Form EIA-860 inventories the generator fleet plant by plant, Electric Power Annual and Electric Power Monthly aggregate generation, sales and price, and SEDS fills the state map. This record is the only one that walks downstream of the meter and describes the unit drawing the power.

That positioning cuts both ways. Freshness belongs to the monthly series, which land about two months behind real time, while the surveys publish on multi-year cycles. But between waves nothing else tells you why the megawatthours moved - which homes added air conditioning, which buildings ran their equipment hours up, which industries switched fuels.

Four things, stated plainly.

First, vintages, not tickers. These are snapshots collected on multi-year cycles - RECS 2020 is the newest complete household wave, with a 2024 initial release underway. Plan trend work as joins across consistent definitions rather than expecting a continuous stream.

Second, geography has a floor. Census division is the standard reporting grain, state detail exists only where disclosure rules allow it, and no addresses survive anywhere. Model within the boundary rather than through it.

Third, weights make the estimates. Every row carries the analysis weight that expands the sample to its population; unweighted averages quietly answer a question nobody asked. Validation rows ship with every delivery so your tabulations reconcile against the published figures.

Fourth, coded columns arrive labeled. Enums like TYPEHUQ and REGIONC come with value-label references attached, so your engineers read words, not integers, on day one.