Methods: location-based emission factor hierarchy analysis

← All insights

Approach

The analysis follows the approach of my 2022 study in Environmental Research Letters: calculate location-based Scope 2 inventories for a large, representative set of buildings using each type of emission factor in the hierarchy, then compare the results.

Buildings and load data. The analysis covers more than 112,000 residential and commercial buildings across 51 balancing areas in the contiguous U.S., the same sample used in my 2022 study. Hourly electricity consumption for each building comes from the 2021 release of the End-Use Load Profiles dataset from the National Laboratory of the Rockies (formerly the National Renewable Energy Laboratory), simulated with typical-year (TMY3) weather. The sample is a stratified random 10% of the buildings of each type in each climate zone in each grid region. Each load profile covers 8,760 hours in local standard time and was aligned to the 2025 calendar.

Emission factors. Hourly production-based and consumption-based CO₂ emission factors for each balancing area come from the U.S. Energy Information Administration’s Hourly Electric Grid Monitor (Form EIA-930) for 2025, with 2024 data used for the prior-year (data vintage) comparisons in Parts 3 and 4. Balancing-area factors represent the operational grid boundary tier of the proposed hierarchy; the other tiers are built by aggregating them (next section).

Inventory calculation. Each building’s inventory is its hourly load multiplied by the relevant emission factor, summed over the year. For monthly and annual factors, the same average value applies to every hour in that month or year.

Emission factor aggregation

Every factor in the analysis is built from the same EIA balancing-area data, so differences between inventories reflect the factor’s spatial, temporal or type characteristics rather than differences between data sources.

Spatial. Balancing areas were aggregated to three coarser boundaries:

How factors are calculated. Every factor in the analysis, at every boundary and time resolution, is calculated the same way: total CO₂ emissions divided by total electricity, over the boundary and period. For consumption-based factors, that is consumed emissions divided by electricity consumed; for production-based factors, emissions from generation divided by generation. Larger boundaries sum emissions and electricity across their member balancing areas each hour, so each balancing area counts in proportion to its size rather than as an equal share.

Temporal. Monthly and annual factors were calculated from the hourly data for each boundary the same way: total emissions over the month or year divided by total electricity over the month or year. They are therefore consumption-weighted (or generation-weighted) averages of the hourly rates, not simple averages.

Measuring error

Every inventory is compared against a benchmark built with the most precise factor available across the U.S.: balancing area, hourly, consumption-based, reporting year (2025).

Error. For each building, error is the percentage difference between its annual inventory and its benchmark inventory. Averages use the absolute value of each building’s error; Part 3’s Figure 1 shows the signed error (positive for an overestimate) and the share of buildings beyond ±5%. Error measures how much a reporter’s inventory would change if it used a less precise factor, not the deviation from “true” emissions, which no dataset observes directly.

Weighting. Regional error rates are weighted by each building’s share of total regional demand. National error rates are further weighted by each balancing area’s share of total national demand. Because errors are absolute values, overestimates and underestimates do not cancel out.

Materiality. Results are compared against the rule of thumb in the GHG Protocol Corporate Standard: an error greater than 5% of an inventory is considered materially misleading.

Near-zero emission rates. Percentage error grows without limit as a building’s benchmark emissions approach zero. A few hydro-dominated balancing areas in the Pacific Northwest, notably the Chelan, Douglas and Grant County public utility districts, have consumption-based rates near zero, so coarser factors overstate their inventories by hundreds or thousands of percent. Their small share of demand limits their weight, but they still raise the averages for coarse boundaries. As a sensitivity check, capping each building’s error at 100% lowers the average error for EIA regions from 18% to about 10%, for interconnections from 29% to about 22%, for the contiguous U.S. from 36% to about 25%, and for the Northwest in Part 2 from 151% to 59%. Balancing-area results, the MISO case study and the other simulated regions are essentially unchanged, and every coarse boundary stays well above the 5% materiality threshold. The posts report uncapped results.

Relation to the 2022 study. This analysis reuses the building sample from my 2022 study, so differences in results come from the emission factors and the error measure: 2025 EIA-930 factors instead of 2019 factors from Singularity’s earlier Carbonara platform, and demand-weighted absolute error instead of signed bias. One balancing area in the 2022 sample (AEC) no longer reports to EIA-930 and is left out, which is why this analysis covers 51 balancing areas rather than 52.

Part 2: operational grid subregions

EIA does not publish emission factors below the balancing-area level, so Part 2 evaluates the operational grid subregion tier (the proposal’s “local boundary”) with two separate case studies. Both use hourly, consumption-based factors, and both treat the subregion’s own factor as the benchmark.

MISO local resource zones. MISO is a single balancing area that spans parts of 15 states, divided into 10 local resource zones (LRZs) that MISO uses for resource adequacy planning. Hourly consumption-based factors for each LRZ and for MISO as a whole come from MISO’s public consumed emissions dashboard, which Singularity built and operates for MISO. The zone factors are for 2025. Each MISO building was assigned to a zone by its state. Michigan, the one state split between two zones, was split by climate zone: buildings in the Upper Peninsula’s climate zone were assigned to Zone 2 and the rest to Zone 7. Because zone boundaries don’t follow state lines everywhere, the assignment is approximate. For each LRZ, inventories calculated with the MISO-wide factor were compared to inventories calculated with that zone’s own factor.

Simulated regionalized grids. Most large U.S. grid operators (MISO, PJM, SPP) combine areas that were once separate balancing areas. To approximate that structure with EIA data, four EIA regions were treated as hypothetical regionalized grids, with their member balancing areas standing in for local zones:

Members are the balancing areas assigned to each EIA region in the 2025 EIA-930 data. The counts include generation-only balancing areas (marked *), because their generation is aggregated into each region’s emission factor. Sample buildings come only from balancing areas that serve load.

• Carolinas: CPLE, CPLW, DUK, SC, SCEG, YAD*
• Florida: FMPP, FPC, FPL, GVL, HST, JEA, SEC, TAL, TEC
• Southwest: AZPS, DEAA*, EPE, HGMA* (reported January–May 2025), PNM, SRP, TEPC, WALC
• Northwest: AVA, AVRN*, BPAT, CHPD, DOPD, GCPD, GRID*, GWA*, IPCO, NEVP, NWMT, PACE, PACW, PGE, PSCO, PSEI, SCL, TPWR, WACM, WAUW, WWA* (reported January–September 2025)

For buildings in each region, inventories calculated with the region-wide factor were compared to inventories calculated with their own balancing area’s factor, weighted by each building’s share of regional demand.

Part 3: data vintage

Prior-year factors. To test the effect of data vintage in Part 3, 2025 building loads were paired with 2024 emission factors, calculated the same way as the 2025 factors, at every spatial, temporal and type combination, and the resulting inventories were compared to the reporting-year benchmark. Prior-year hourly factors were matched to 2025 hours by calendar date and hour, so 2 p.m. on March 3, 2025 is paired with 2 p.m. on March 3, 2024, whatever the day of the week. Prior-year monthly and annual factors were matched by month and by year.

Reporting-year data as the benchmark. Like the rest of the series, this analysis takes reporting-year EIA-930 factors at face value. Whether near-real-time factors, which are based on fleet-level averages, are reliable enough compared with bottom-up datasets such as Open Grid Emissions (OGE), which are built from plant-level data but published 10–12 months after the end of each year, will be tested in a separate post outside this series. Singularity publishes both OGE and a near-real-time dataset (the Grid Carbon API).

Part 4: ranking the hierarchy

Ranking every combination. For Part 4, inventory error was calculated for every combination of spatial boundary (balancing area, interconnection, and the contiguous U.S. for reference), type (consumed, produced), data vintage (reporting year, prior year) and temporal granularity (hourly, monthly, annual). Part 4 shows prior-year factors as annual averages only, consistent with the proposal’s rule for older data, which gives 24 inventories per building. The tiers were then ranked from lowest to highest demand-weighted average error, and the ordering of the four qualities was read from that ranking.

Mapping real datasets to tiers. Each publicly available U.S. dataset was placed at the highest tier it can serve for a reporter closing a 2025 inventory:

• MISO local resource zone factors: operational grid subregion, consumed, reporting-year hourly.
• Electricity Maps, Singularity’s Grid Carbon API, EIA and WattTime: balancing area, consumed, reporting-year hourly.
• Open Grid Emissions: balancing area, consumed, prior year. It publishes hourly, monthly and annual factors, but enters as an annual factor because prior-year factors are limited to annual averages.
• eGRID: balancing area, produced, prior-year annual.
• IEA: national, produced, prior-year annual.

Only these five tiers have a dataset. The five were then ranked under three orderings of the four qualities, to test whether the ordering changes which dataset a reporter would use: the public consultation proposal (vintage, spatial, temporal, type); the same order with type and temporal switched (vintage, spatial, type, temporal); and the error-minimizing order (spatial, type, vintage, temporal).

Limitations

• Geography: the analysis covers the contiguous U.S. only. U.S. grids are operationally diverse, so the results may be reasonably representative of other locations, but they may not generalize globally.
• One year: results are for 2025 inventories, with 2024 as the prior year. The relative importance of each quality may differ in other years, especially as grids change faster.
• Weather: building loads are simulated with typical-year weather rather than 2025 weather, so the analysis doesn’t capture weather-driven links between when buildings use electricity and how clean the grid is in those hours. This likely understates the value of hourly factors.
• Loads: residential and commercial buildings only. Industrial loads, which often run around the clock, may produce different results.
• Data quality: EIA balancing-area data is self-reported by grid operators and has known gaps and errors.
• Storage: EIA’s consumption-based factors account for electricity imports and exports, but not for electricity (and its emissions) shifted through time by energy storage.
• Operational grid subregions: EIA publishes no subregional factors, so this tier is tested only through the case studies in Part 2.
• Benchmark: error is measured against the most precise broadly available factor, not against true emissions.