5 Methods and data sources
Electricity emissions are estimated by collecting reported electricity deliveries from all utilities operating within our study area and applying EPA eGRID emissions factors to convert megawatt-hours delivered into metric tons CO2e. The inventory takes a demand-side approach: it accounts for electricity delivered to end-use customers (residential, commercial, and industrial) and does not include transmission and distribution losses or electricity resold to utilities outside the study area. All emission factors are sourced from the EPA eGRID database for the MROW subregion.
6 County-level estimates
6.1 Source data
County-level electricity deliveries are derived from state regulatory filings in Minnesota and a combination of state and federal filings in Wisconsin, both of the highest quality rank (Table B.2).
6.1.1 Minnesota
Under Minnesota Administrative Rules Chapter 7610 (Minnesota Department of Commerce 2005), all electric utilities authorized to do business in Minnesota must file annual reports of county-level energy deliveries in megawatt-hours (MWh). These filings provide the primary source of measured utility-by-county-by-year electricity data.
The ingestion pipeline parses county-level delivery data from the “ITS DELIVERIES TO ULTIMATE CONSUMERS BY COUNTY” section of each utility’s annual report workbook using anchor-based detection (locating the COUNTY/MWH header row programmatically rather than relying on fixed cell positions). Utility names are harmonized through a lookup table to resolve inconsistencies across filing years (e.g., “Northern States Power Company” and “Xcel Energy” map to the same entity).
Great River Energy and member cooperatives. Great River Energy (GRE) is a wholesale power cooperative that supplies electricity to 28 member cooperatives. Because GRE’s 7610 filing reports all member co-op retail deliveries combined, the individual member cooperatives’ filings are excluded to avoid double counting. In Anoka County, where both GRE and Connexus Energy (a GRE member) file independently, GRE’s filing is dropped for that county. GRE did not file 7610 reports for 2013–2015, and their 2022–2023 filings were flagged as suspect (implausibly low, likely due to member cooperatives beginning to file independently); these years are filled from EIA 861 data as described below.
Gap-filling and backcast. For utility-county-years where 7610 filings are missing (including GRE’s gap years and all years prior to the earliest available filing), deliveries are estimated using EIA Form 861 data. The EIA 861 annual report provides statewide electricity sales by utility; county-level estimates are derived by applying each utility’s county delivery proportions from the nearest available 7610 filing to the EIA 861 statewide total. Gap-fill and backcast use the same method — a unified fill loop covers all missing years from 2005 through the latest available filing year for each utility.
6.1.2 Wisconsin
Wisconsin’s two study area counties (St. Croix and Pierce) are served by seven electric utilities. Under Wis. Stat. § 196.07, investor-owned and municipally-owned utilities must submit annual reports to the Public Service Commission that include total energy deliveries and county-level customer counts. For the four cooperative utilities that do not file state reports, we rely on EIA Form 861 data.
Because Wisconsin utilities do not report deliveries at the county level, county allocations are estimated. For utilities that report county-level customer counts, county deliveries are estimated by multiplying statewide deliveries by the proportion of the utility’s customers in each county:
\[ \text{MWh}_{i,j} = \text{MWh}_i \times \frac{\text{Customers}_{i,j}}{\text{Customers}_{i,\text{total}}} \]
For cooperative utilities without county-level customer data, county deliveries are allocated by population share within each utility’s service area, determined by spatially joining Census block centroids to utility service area polygons.
6.2 Sector decomposition
County-level electricity totals from 7610 filings and Wisconsin filings represent aggregate deliveries to all customer classes. To partition these totals into residential, commercial, and industrial sectors, we apply the following decomposition:
\[ \text{County total} = \text{Residential} + \text{Non-residential remainder} \]
\[ \text{Non-residential remainder} = \text{Commercial} + \text{Industrial} \]
Each component is estimated as follows:
- Residential: the sum of CTU-level residential electricity from the random forest prediction model (described below), aggregated to county.
- Commercial and Industrial: the non-residential remainder (county total minus aggregated CTU residential) is split into commercial and industrial using NREL SLOPE modeled ratios. NREL SLOPE provides county-level forecasts of electricity consumption by sector (Ma et al. 2019; NREL 2017); the commercial-to-industrial ratio from NREL is averaged across available years and applied to the non-residential remainder.
This approach parallels the natural gas sector decomposition but is simpler: electricity does not require separate power plant, industrial combustion, or refinery subtractions because those categories are not present in the delivery totals (power plant self-consumption is not reported as a retail delivery, and large industrial loads are already included in the aggregate).
6.3 NREL SLOPE sector comparison
The chart below compares two approaches to sector decomposition for the seven-county metro:
- CTU model (current method): Residential electricity is estimated from CTU-level random forest predictions aggregated to county. The non-residential remainder (county total minus residential) is split into commercial and industrial using NREL SLOPE modeled ratios.
- Pure NREL proportioning (prior method): All three sector shares (residential, commercial, industrial) are derived from NREL SLOPE modeled proportions applied directly to county utility delivery totals. This remains the production method for the four collar counties (Chisago, Sherburne, St. Croix, Pierce) where CTU-level modeling is not available.
The two methods share the same county-level electricity totals and the same NREL commercial-to-industrial ratio. The difference lies entirely in how residential is estimated: the CTU model anchors residential to utility-reported and modeled CTU data, while pure NREL proportioning applies NREL’s forecasted residential share (built from a 2016 base year) to the county total. This difference in residential magnitude cascades into the commercial and industrial estimates, since both are derived from the remaining non-residential balance.
The CTU model demonstrates that the NREL model appears to underestimate residential electricity shares across the region. Furthermore, some uncertainty remains on proportioning between industrial and commercial electricity demand, as Washington and Dakota counties in particular should demonstrate outsized industrial electricity consumption based on large point sources. As it stands, they do appear to have a closer ratio of commercial to industrial. Ultimately, precise industrial point source data is hidden behind privacy laws and a more accurate estimation is unavailable, to the best of our knowledge.
6.4 Emissions calculation
County emissions are calculated by applying eGRID MROW subregion emissions factors to MWh delivered:
\[ \text{Emissions}_{\text{county,sector}} = \text{MWh delivered} \times \text{Emission Factor (MT CO}_2\text{e / MWh)} \]
The eGRID emissions factor is subregion-specific, reflecting the generation mix in the Midwest Reliability Organization West region. It includes CO2, CH4, and N2O components, converted to CO2e using IPCC AR5 100-year global warming potentials. Importantly, the eGRID factor changes annually as the grid mix evolves — coal retirements and wind additions reduce the factor over time, which drives much of the observed decline in electricity emissions since 2005.
This inventory uses a location-based accounting approach per the GPC protocol: emissions are attributed based on the average grid mix of the subregion where electricity is consumed, without adjusting for renewable energy certificates (RECs) or other market-based instruments.
7 CTU-level estimates
7.1 Source data
CTU-level (city, township, and unorganized territory) electricity data comes from multiple utility sources, compiled and prioritized in the following order:
- Xcel Energy community energy reports. Xcel publishes annual Community Energy Reports documenting electricity deliveries to specific communities in their service area (2015–2023). These provide residential and business MWh by CTU for cities served by Xcel. Because Xcel reports a combined “Business” category rather than separate commercial and industrial figures, NREL SLOPE modeled ratios are used to disaggregate the non-residential component at the county level.
- Dakota Electric Association. Direct data request from Dakota Electric for CTU-level residential and business deliveries, covering communities in Dakota and Scott counties within their service area. An interpolation step addresses a 2018 data gap using 2017 and 2019 anchor values.
- Connexus Energy. CTU-level data from Connexus covering communities in Anoka County. A quality control step identifies and addresses CTU-level allocation anomalies (e.g., over-reported or under-reported deliveries for specific cities) by nulling suspect city-years so they fall through to modeled estimates.
- Municipal utilities. Where available, data from municipal electric utilities (e.g., City of Chaska Electric, Princeton Public Utilities, Shakopee Public Utilities, North Branch) compiled from 7610 filings.
- Prior Met Council SQL records and RII data. Historical CTU-level electricity data from earlier Met Council inventory work and the Regional Indicators Initiative. These are used as fallback sources and are overwritten by any recent utility requests from Met Council.
A data quality pipeline identifies and addresses anomalies before modeling, including year-over-year spike detection for utility data with structural breaks, de minimis utility filtering, and unit conversion corrections.
7.2 Random forest model
For CTUs and years lacking direct utility data, residential and business electricity consumption are predicted separately using random forest models. The models are trained on CTUs with known utility data (approximately 2010–2023). Predictors include:
- UrbanSim housing unit or job projections
- NOAA heating degree days
- Metropolitan Council thrive designation (community classification)
- Parcel-derived building stock characteristics (2021)
For CTUs with at least some years of known data, a mean-scale correction is applied: the ratio between the mean observed utility value and the mean predicted value across all available years is used to multiplicatively adjust all predicted years. This approach anchors predictions to the observed level for each CTU while allowing the random forest to capture temporal trends. CTUs with no utility data in any year receive pure random forest predictions. Cities where the primary utility provider never supplied CTU-level data (e.g., Champlin, Columbus) are excluded from the model and receive county-proportional estimates.
The models are limited to 2010 onward by UrbanSim data availability.
7.3 Pre-2010 backcast
For years prior to 2010, CTU electricity consumption is estimated by applying each CTU’s earliest three years of sector proportions to the county-level MWh totals. This approach parallels the natural gas backcast and preserves the relative distribution of electricity across CTUs within each county while anchoring to the county-level totals from the 7610/EIA pipeline.
8 Validation
8.1 County totals: CTU model vs. county activity
The chart below compares CTU-level modeled electricity (aggregated to county) against county-level delivery totals from 7610 filings and EIA 861 gap-fills. Similar to observations in natural gas modeling, we see under-estimates in Dakota and Washington counties, where large point source industrial facilities are not reported in CTU data due to privacy laws. The noticeable underestimate in Scott County is less clear, though may also be masked industrial data.
8.2 Correlation with population
We would expect counties with higher population to have higher electricity emissions.
8.3 Assumptions and limitations
Demand-side accounting. This inventory counts electricity delivered to end-use customers. It does not include transmission and distribution losses (estimated at approximately 5% nationally by EIA) or electricity generated within the region and exported. This is consistent with GPC protocol for community-scale inventories.
Location-based emissions factors. Using the eGRID MROW subregion factor assigns the same emissions rate to all electricity consumed in the region, regardless of whether a specific utility has a cleaner or dirtier generation portfolio. A market-based approach (adjusting for RECs and utility-specific generation) would yield different county-level estimates, particularly for utilities with significant renewable procurement.
Sector decomposition. The commercial-industrial split relies on NREL SLOPE modeled ratios rather than direct utility reporting by sector. While Xcel community reports provide a combined “Business” category, no utility in the region reports separate commercial and industrial deliveries at the county level. The NREL ratios are derived from a 2016 base year forecast and may not perfectly reflect the actual commercial-industrial mix in any given year.
CTU random forest model. As with natural gas, the RF model is limited to 2010 onward by UrbanSim data availability. Pre-2010 CTU estimates use proportional allocation from county totals. The mean-scale correction assumes that the ratio between observed and predicted values is stable over time; for communities with rapid growth or structural changes in electricity demand, this assumption may introduce bias. A systematic shift in Xcel’s application of the 15/15 privacy rule between 2021 and 2022 affects the business electricity data for approximately 15 cities; this is under investigation and may be revised pending utility response.
Wisconsin county allocation. Wisconsin county-level estimates rely on customer count or population proportions to allocate utility-wide deliveries to individual counties. This implicitly assumes uniform per-account or per-capita demand within each utility’s service area, which may not hold where residential and commercial/industrial mixes differ across counties.
GRE member cooperative reporting. The inventory uses GRE’s consolidated filing as a substitute for individual member cooperative filings. This may represent a marginal undercount for cooperatives that source small amounts of electricity from non-GRE suppliers. For 2022–2023, GRE’s own 7610 filings appeared suspect (likely due to member cooperatives beginning to file independently) and were replaced with EIA 861 estimates.