Electricity consumption variation versus economic structure during COVID-19 on metropolitan statistical areas in the US

COVID-19 spread rapidly in the U.S. in early 2020, prompting stay-at-home orders, widespread work-from-home practices, and reduced mobility. These shifts altered energy demand patterns across fuels and electricity, with notable regional differences. Prior observations showed increased renewable generation shares and reductions or shifts in electricity demand peaks. Because electricity consumption (EC) correlates with economic activity, and is also shaped by economic structure, understanding how the composition of metropolitan economies influenced EC changes during the initial pandemic months is important. This study takes each metropolitan statistical area (MSA) as the analysis unit to estimate EC changes between April–May 2019 and April–May 2020 for both total and residential sectors. It aggregates county-level estimates to 380 MSAs across the continental U.S. (covering 86% of the population and 87% of EC), aiming to reveal how economic structure and COVID-19 incidence relate to EC variation in the early, unprepared phase of the pandemic.

Prior work and reports have documented: (1) increases in renewable generation shares during COVID-19 due to policy support and declining costs despite supply chain delays; (2) overall reductions in electricity demand and shifts in load shapes, including lower daily peaks occurring later; (3) heterogeneous regional impacts across U.S. system operators (e.g., demand reductions in the Midcontinent region versus relatively stable demand in Florida); and (4) relationships between EC and macroeconomic indicators such as GDP, while also noting that EC depends on economic structure, not only output, implying that structural changes can shift EC. Studies and operator reports (e.g., EPRI, EIA, NYISO, MISO, PJM, CEC) provide partial regional evidence of these patterns but lack a comprehensive, nationwide metropolitan-level view. This work addresses that gap by linking metropolitan economic composition to EC changes during the initial pandemic months.

Study scope and period: The analysis covers 380 MSAs in the continental U.S. (excluding 4 MSAs in AK/HI) and compares April–May 2019 versus April–May 2020 to capture the initial lockdown period.

Data sources: County/MSA/state GDP (BEA), population (U.S. Census), COVID-19 county-level cases (JHU aggregated to MSA), and state-level electricity consumption (EIA). Information technology (IT) intensity by industry was used for GDP growth adjustment. Data processing used NumPy, pandas; clustering/statistics used scikit-learn and SciPy.

COVID-19 incidence: Daily new confirmed cases per 100,000 population were computed using a 7-day moving average, then classified by thresholds: Low (0), Medium [1,10), High [10,25), Critical [25,∞). For each MSA, the most frequent level during April 1–May 31, 2020 was assigned as its incidence level.

Economic structure: 2019 MSA-level GDP by industry (35 categories reduced to 20 representative categories to mitigate overlap) characterized pre-pandemic economic composition. Missing values (due to BEA suppression) were imputed using recent-year averages where possible; GDP shares were normalized (0–1), categories with entirely missing values were backfilled proportionally, and a fifth-root transform alleviated right skewness. K-means clustering (Euclidean distance, elbow method for k) grouped MSAs within incidence strata into clusters reflecting dominant economic structures. For interpretability, categories were regrouped: management/administrative/educational as MAE services; information/finance-insurance/professional as high-end services; and an ‘Other’ category combining construction, wholesale, retail, accommodation/food, arts/entertainment/recreation, and other services.

Electricity consumption estimation: Because MSA-level EC is not directly available, a two-step estimation was implemented.

1. County-level EC allocation from state totals: Guided by observed linear relations at the county level (California 2019) between total EC and GDP, and between residential EC and population, state-level EC was proportionally allocated to counties within each state:

• Total EC for county c: ECT_c = (GDP_c / GDP_s) × ECT_s
• Residential EC for county c via equal per-capita within-state allocation: ECR_c / P_c = ECR_s / P_s Here, GDP_c is county GDP (current dollars), GDP_s is state GDP (annualized quarterly, current dollars), ECT_s/ECR_s are state-level total/residential EC; P_c/P_s are county/state populations. County estimates were aggregated to MSAs for 2019 and 2020, then percentage changes were computed for total and residential EC.

2. Estimating 2020 county GDP: 2019 county GDP was grown to 2020 using state real GDP chain-type quantity indices (Q2 2019 to Q2 2020) to represent April–May growth, with adjustments for industry IT-intensity to reflect varying ability to work remotely. This produced county-level 2020 GDP inputs for the proportional EC allocation. Details of IT-intensity adjustments and growth-rate derivation are provided in the Methods (including formulas referencing state-level Q2 indices and industry digital worker shares).

Statistical analysis: EC variation distributions across economic clusters were compared using two-sided Wilcoxon rank-sum tests (each cluster vs. all MSAs) for both total and residential EC. Spatial patterns were summarized by state/region with 95% confidence intervals (CIs). A sensitivity analysis examined robustness to updated COVID, GDP, and EIA EC data; key clustering results and most statistical conclusions remained stable.

Verification: Estimated metropolitan-level patterns were cross-checked against operator/regulator reports: CEC (California), MISO, and PJM. Despite differences in territory/time windows (weekday vs. all days), the magnitudes of reductions/increases were broadly consistent, supporting credibility of the estimates.

Overall patterns:

• Total EC decreased across most MSAs during April–May 2020 versus April–May 2019, while residential EC generally increased.

Spatial heterogeneity in total EC change:

• Largest total EC decline: Muskegon, MI (−15.18%). Other Michigan MSAs: >12% declines; statewide mean decline 95% CI [−14.67%, −13.89%], n=15.
• Midwest (IL, IN, IA, KS, MI, MN, MO, NE, ND, OH, SD, WI): mean decline ≈ 8.88%, 95% CI [−9.71%, −8.06%], n=96.
• Northeast (CT, ME, MA, NH, NJ, NY, PA, RI, VT): ≈ −7.45%, 95% CI [−7.89%, −7.01%], n=51.
• West Coast: OR [−2.13%, −1.33%], n=8; CA [−7.13%, −6.63%], n=26; WA [−7.08%, −4.04%], n=13.
• Southeast and nearby (AL, FL, GA, KY, MD, MS, NC, SC, TN, VA, WV): [−9.69%, −8.07%], n=114; Florida had smaller declines [−3.65%, −3.17%], n=22.
• Increases: LA [+0.69%, +1.21%], n=9; TX [+1.32%, +2.14%], n=25; NM [+2.70%, +3.51%], n=4; AZ [+8.59%, +9.85%], n=7 (largest MSA increase: Sierra Vista–Douglas, AZ +10.47%). Some MSAs in ND, ID, NV also rose slightly.

Spatial heterogeneity in residential EC change:

• Nationally, residential EC increased. Largest increase: Phoenix–Mesa–Chandler, AZ (+29.05%).
• AZ and NV: >20% increases; 95% CIs AZ [26.88%, 28.65%], n=7; NV [24.30%, 25.92%], n=3.
• NM [14.25%, 17.23%], n=4; TX [8.83%, 9.95%], n=25; LA [7.23%, 8.04%], n=9; FL [7.53%, 9.23%], n=22.
• Northeast: [8.32%, 10.18%], n=51; CA [9.71%, 10.73%], n=26; OR [7.61%, 9.38%], n=8; WA [3.40%, 4.88%], n=13.
• Central states (UT, CO, KS, OK, MO, IL, IN, KY): aggregated [5.93%, 7.80%], n=60.
• Exceptions: Slight residential decreases in parts of the Southeast—NC, SC, GA, AL [−5.25%, −4.42%], n=51; largest drop: Florence, SC (−8.77%). Virginia showed near-flat to slight decreases [−1.62%, −0.81%], n=7.

Economic structure associations (clusters):

• Mining share: MSAs with higher mining industry shares exhibited significantly smaller reductions in total EC than other MSAs (Wilcoxon rank-sum comparing mining-heavy clusters vs. others: p < 1e−4; clusters II and IV showed significant differences vs. all MSAs in Fig. 4).
• Agriculture/forestry and manufacturing: Clusters with higher shares (V, VI) tended to have greater total EC declines than average; Cluster VI’s decline was less severe than Cluster V, plausibly due to a higher mining share in VI.
• High-end and MAE services (information, finance/insurance, professional; management/administrative/educational): Total EC reductions not significantly different from other MSAs (Wilcoxon p=0.3919 combining relevant clusters; individual cluster tests largely nonsignificant).
• Real estate/leasing and public administration: Clusters with higher shares (I, VIII) showed statistically smaller total EC reductions (combined Wilcoxon p=0.0452, n1=49, n2=331, W=2.0032).
• COVID-19 incidence level: Total EC changes did not show a clear pattern across incidence levels; residential changes were broadly similar across incidence categories.

Residential EC by economic structure:

• Residential EC increased across all clusters. Cluster IV showed a higher-than-average increase; clusters VII and VIII appeared higher but were not statistically significant due to small sample sizes. Median residential EC increases across clusters were roughly 7–10%.

External verification:

• California: CEC reported ~9% weekday total EC reduction in April 2020 vs April 2019; MSA-level estimate for CA was −6.9% (95% CI [−7.1%, −6.6%], n=26) over April–May. CEC reported 8.9–12.4% residential increase (Jan–May 2020 vs 2019); MSA estimate for April–May was ~10.2%.
• MISO footprint: Operator observed −9.34% total EC (April–May 2020 vs 2019); MSA estimate across states in MISO territory was −8.0% (95% CI [−9.1%, −6.9%], n=84).
• PJM footprint: Operator reported ~10–14% decrease in early May 2020 and 6–11% later May–early June; MSA estimate for April–May was −9.3% (95% CI [−10.1%, −8.8%], n=118). These alignments support credibility of the metropolitan estimates.

The study addresses how the initial COVID-19 lockdowns affected electricity use at the metropolitan scale and how economic structure shaped these impacts. By constructing county-to-MSA EC estimates and clustering MSAs by pre-pandemic economic composition, the analysis demonstrates that total EC reductions varied with industry shares: mining and real estate/public administration-oriented MSAs experienced smaller declines, while agriculture/forestry and manufacturing-oriented MSAs saw larger declines. Intelligence-intensive service economies showed reductions similar to overall averages, likely reflecting shifts of computing loads to homes offset by reduced commercial HVAC and lighting. Residential EC increases were widespread and largely insensitive to economic structure or incidence level, consistent with work-from-home and stay-at-home orders. These findings can inform grid planning and operations, suggesting that future “pandemic-ready” or crisis-aware operational constraints and resource planning should consider local economic composition to anticipate demand shifts. The dynamic correlation observed between COVID-19 incidence/deaths and EC over time implies EC can serve as an ancillary indicator of public-health-driven societal activity changes. Given the possibility of longer-term shifts in work practices, some EC pattern changes may persist beyond the emergency period, with implications for load forecasting, rate design, and infrastructure maintenance scheduling.

This paper provides the first nationwide metropolitan-level estimates of EC changes during the initial months of COVID-19 and links them to local economic structures. Key contributions include: (1) a practical method to estimate MSA-level total and residential EC from state-level EC using county GDP/population shares; (2) clustering-based characterization of metropolitan economic structures; and (3) evidence that total EC reductions depended on industry composition (smaller declines for mining, real estate/public administration; larger declines for agriculture/forestry and manufacturing), while residential EC generally increased irrespective of structure or incidence. Consistency with operator/regulator reports supports the estimates’ credibility. Future work should incorporate additional data sources (e.g., climate variables like degree-days, GDP per capita), refine modeling beyond proportional allocation (e.g., multivariate/panel models), validate county-level relationships across more states as data become available, and examine temporal dynamics of EC–incidence correlations to improve demand forecasting and crisis-responsive grid operations.

• The EC estimation relies on extrapolating linear relationships observed in California (county-level total EC vs GDP; residential EC vs population) to counties nationwide. Although state-level linearity supports aggregation, county-level validity in other states remains unconfirmed due to data scarcity. Degraded linearity elsewhere could introduce uncertainty; data transformations may help.
• The proportional allocation method is an intentionally simple approach. Accuracy could be improved by incorporating more detailed variables (e.g., climate/degree-days, GDP per capita, additional economic indicators) and more sophisticated modeling frameworks. Broader energy supply/demand assessments could leverage panel data across electricity, petroleum, and gas to better capture cross-sector interactions.
• Source datasets (COVID-19 cases, GDP, state EC) are periodically updated. Sensitivity analyses showed small quantitative shifts and one change in residential cluster significance (Cluster V from significant to nonsignificant), but overall pattern conclusions remained robust.