Source-linked AI summary
A Cross-Domain Approach to Analyzing the Short-Run Impact of COVID-19 on the U.S. Electricity Sector
Guangchun Ruan, Dongqi Wu, Xiangtian Zheng, Haiwang Zhong, Chongqing Kang, Munther A. Dahleh, S. Sivaranjani, Le Xie
TL;DR
Existing research lacked consistent, weather-calibrated, cross-domain evidence on COVID-19’s impact on U.S. electricity consumption. This paper integrates electricity-market, public-health, weather, mobility, and satellite data to assess that impact, finding consumption reductions across markets ranging from 6.36%–10.24% in April and 4.44%–10.71% in May.
Problem
Existing assessments lacked consistent comparisons, rigorously calibrated no-pandemic baselines, and integration of public-health and social-distancing data.
Method
The study develops COVID-EMDA+, an open-access cross-domain data hub and statistical models integrating electricity-market, health, weather, mobility, and satellite data.
Results
Electricity consumption declined across all markets, by 6.36%–10.24% in April and 4.44%–10.71% in May.
Takeaways & Limitations
The analysis can inform short-term grid planning, load forecasting, and policy assessment by identifying retail mobility as a key influence on consumption changes.
Takeaways & Limitations
Satellite imagery was used only for visualization, not numerical analysis, because low sampling frequency and cloud contamination limited valid observations.
Abstract
from arXiv · showhide
The novel coronavirus disease (COVID-19) has rapidly spread around the globe in 2020, with the U.S. becoming the epicenter of COVID-19 cases since late March. As the U.S. begins to gradually resume economic activity, it is imperative for policymakers and power system operators to take a scientific approach to understanding and predicting the impact on the electricity sector. Here, we release a first-of-its-kind cross-domain open-access data hub, integrating data from across all existing U.S. wholesale electricity markets with COVID-19 case, weather, cellular location, and satellite imaging data. Leveraging cross-domain insights from public health and mobility data, we uncover a significant reduction in electricity consumption across that is strongly correlated with the rise in the number of COVID-19 cases, degree of social distancing, and level of commercial activity.
Introduction
The introduction highlights uncertainty and gaps in assessing COVID-19’s electricity-sector impacts, then presents COVID-EMDA+ as a cross-domain open-access data hub for tracking and measuring those impacts in the U.S.
- Motivation: Uncertainty remains about the duration and severity of COVID-19’s impact on the electricity sector as the U.S. responds and states reopen the economy.The introduction notes that scholarly research on this impact remains limited.
- Research gaps: Existing assessments lack consistent criteria, limiting comparability across geographical locations.The introduction identifies this as the first major gap in prior research.
- Research gaps: Existing statistical analyses often do not rigorously calibrate baseline electricity consumption without the pandemic while accounting for exogenous factors such as weather.The introduction identifies baseline calibration and weather effects as additional methodological gaps.
- Contribution: COVID-EMDA+ integrates electricity-market data with COVID-19 public-health, weather, mobile-location, and satellite-imagery data.The hub is intended to track and measure COVID-19’s impact on the U.S. electricity sector.
Cross-domain Data Hub: COVID-EMDA+
COVID-EMDA+ is a public cross-domain data hub integrating U.S. electricity-market, weather, public-health, mobility, and satellite-imaging data. It supports visualizing and quantifying COVID-19-related electricity-consumption reductions and their correlations with cases, social distancing, and commercial activity.
- Data Hub Development: COVID-EMDA+ integrates electricity-market, weather, mobile-device location, and satellite-imaging data into a single ready-to-use open-access format.The hub is publicly available on GitHub.
- Electricity-Market Data: Across seven U.S. regional electricity markets, the hub aggregates load, generation mix, and day-ahead locational marginal price data.The markets are CAISO, MISO, ISO-NE, NYISO, PJM, SPP, and ERCOT.
- Public Health and Mobility Data: The hub combines COVID-19 case data with county-level social-distancing measures and visits to points of interest from mobile-device location data.The mobility data include visits to categories such as restaurants and grocery stores.
- Satellite-Imaging Visualization: Reduced night-time-light brightness provides a visual representation of COVID-19’s effect on electricity consumption in New York City.The hub uses NASA’s VNP46A1 “Black-Marble” satellite imagery dataset for this visualization.
- Cross-Domain Insights: COVID-EMDA+ enables analysis of electricity-consumption reductions correlated with COVID-19 cases, social distancing, and commercial activity.Retail mobility is defined as the number of daily visits to retail establishments and is used as an indicator of commercial activity.
Impact of Public Health, Social Distancing, and Commercial Activity on Electricity Consumption During COVID-19
Electricity consumption changes are dynamically linked to public health, social distancing, and commercial activity, with retail mobility emerging as the most significant and robust factor across cities. COVID-19 cases show a weaker direct influence, while sensitivity can remain high even where overall consumption reductions are mild.
- Temporal dynamics: The influencing factors have different temporal dynamics: retail mobility began declining in late February 2020 and continued falling through late April in New York City.Electricity consumption, public health, stay-at-home, work-on-site, and retail-mobility indicators changed on widely varying time scales.
- Key findings: Retail mobility is the most significant and robust factor influencing electricity-consumption decreases across all cities.It explains a significant share of consumption changes in both variance-decomposition and impulse-response analyses.
- Key findings: A 1% decrease in retail-sector mobility produces a 0.78% reduction in Houston electricity consumption.
- Key findings: New confirmed COVID-19 cases are not a strong direct influence on electricity-consumption changes across cities.Impulse-response analyses show low consumption sensitivity to this factor, despite its accessibility as an indicator.
- Key findings: High sensitivity to an influencing factor can occur in cities with only mild overall electricity-consumption reductions.Houston is highly sensitive to retail-mobility variations despite its relatively modest consumption change.
Discussion
The study introduces an open-access cross-domain data hub for analyzing COVID-19’s impact on the U.S. electricity sector and demonstrates its use in an initial assessment. The analysis supports short-term grid planning and policy analysis while motivating future integration of socioeconomic data.
- Contribution: The study introduces an easy-to-use open-access data hub aggregating multiple data sources to track and analyze COVID-19’s impact on the U.S. electricity sector.The hub enables cross-domain analysis during and after the pandemic and supports initial quantification of the impact’s intensity and dynamics.
- Implications: The cross-domain analysis can inform power-system operators’ short-term planning, grid operation, load forecasting, and quantitative assessments of renewable-energy curtailment.
- Future research: Future research will integrate demographic data and the social vulnerability index into the COVID-EMDA+ data hub to study COVID-19’s increased energy burden on vulnerable populations.
Experimental Procedures
The study integrates electricity-market, weather, public-health, satellite, and mobile-location data, then harmonizes these sources and models pandemic-related consumption reductions against an estimated no-pandemic baseline.
- Data Integration: The COVID-EMDA+ hub integrates data from all U.S. electricity markets with weather, COVID-19 health, satellite-imagery, and mobile-device location data.The integration is designed to obtain cross-domain insights into COVID-19’s impact on the electricity sector.
- Data Processing: A processing flowchart reorganizes heterogeneous sources around data consistency, compaction, and quality control.Market-specific parsers standardize files into indexed tables, geocoding aligns geographic scales, and quality controls address redundancy, missingness, duplicates, and outliers.
- Backcast Model: The ensemble backcast estimates electricity consumption without the pandemic, allowing differences from actual metered consumption to quantify pandemic impact.Inputs include calendar, weather, and economic factors, with weather represented through selected quantiles and GDP growth included as an economic input.
- Backcast Model: A four-layer fully connected ReLU neural network was selected as the base model for accuracy and robustness, with candidates trained using 85% of 2018 and 2020 data.Models are validated on the remaining 15% using the L2 norm of monthly prediction errors.
- Reduction Model: A restricted VAR models electricity-consumption reductions using selected COVID-19, stay-at-home, dwell-time, worker, and retail-mobility variables.The model captures linear correlations among multiple time series through lagged variables.
Introduction
The supplement organizes supporting materials into additional figures, cross-domain dataset notes, and statistical methods for the COVID-EMDA+ analysis. It covers the data hub, mobility and electricity visualizations, dataset definitions, restricted VAR procedures, and statistical test results.
- Supplementary Methods: The supplementary methods describe pre-estimation preparation, restricted VAR estimation and verification, post-estimation analysis, and model selection.The materials provide the procedural basis for establishing and evaluating the statistical analysis.
- Supplementary Figures: The supplementary figures cover the COVID-EMDA+ data hub architecture and visualizations of night-time lights, stay-at-home populations, points-of-interest visits, COVID-19 cases, and electricity consumption.Additional figures also present VAR results, data quality control, Google Trends activity, and cross-domain variable transitions.
- Supplementary Notes: The supplementary notes define the night-time-light, mobile-device-location, retail, and COVID-EMDA+ data sources used in the analysis.They also discuss restricted VAR parameter choices comparing COVID-19 cases with hospitalizations and deaths.
- Supplementary Tables: Supplementary tables report restricted VAR parameters, restricted VAR statistical test results, COVID-19 metric availability and correlations, and tests using different COVID-19 indicators.These tables complement the documented model procedures and dataset definitions.
Supplementary Figures · Supplementary Figure S-1: Architecture of COVID-EMDA+ data hub
The supplementary material describes COVID-EMDA+ as a cross-domain data hub integrating electricity, weather, public-health, mobility, and satellite-imagery data across dates, data types, and locations. Supplementary analyses visualize COVID-19-related changes in night-time lights, mobility, electricity consumption, and data quality, while noting limitations of satellite and mobile-device data.
- Supplementary Figure S-1: Architecture of COVID-EMDA+ data hub: COVID-EMDA+ cross-references electricity-market, weather, public-health, mobile-location, and satellite-imagery data across dates, data types, and RTO or city locations.Geocoding coordinates the five sources geographically, while processing targets data consistency, compaction, and checking.
- Supplementary Figure S-3: Comparison of Stay-at-home Population: Stay-at-home population increased during the pandemic, with regional differences associated with the stringency and effectiveness of stay-at-home policies.The comparison uses February 12 as the baseline and selected Wednesdays that were not holidays.
- Supplementary Figure S-4: Change of Visits to Points of Interest (POIs): Visits to restaurants, groceries, and health-and-personal-care locations declined suddenly across four hotspot cities beginning March 13, with similar patterns and regional divergences.Daily visit counts are normalized to February 15 as the baseline and span February 15 to April 25, 2020.
- Supplementary Figure S-4: Change of Visits to Points of Interest (POIs): Supplementary traces compare electricity-consumption reductions with new COVID-19 cases and stay-at-home populations across multiple U.S. cities.The visualizations include Boston, Houston, Kansas City, New York City, Philadelphia, Chicago, and Los Angeles; cases and stay-at-home populations use weekly moving averages.
- Supplementary Figures: The data hub supplements cross-domain analysis with restricted VAR results, quality-control procedures, Google Trends activity, and seasonal trend-transition visualizations, while SafeGraph coverage represents about 10% of U.S. devices.Quality control includes outlier detection and missing-data recovery using backup data and historical trends.
- Supplementary Note SN-1: Description of the Night-Time Light Dataset: Satellite night-time-light images were used only for visualization, not numerical analysis, because sampling was once daily and many observations were cloud-contaminated.The dataset therefore provides visual evidence rather than a quantitative electricity-consumption measure.
- Supplementary Figure S-2: Night Time Light Images in COVID-19 Hotspot Cities: Night-time light brightness visibly declined in Boston, New York City, Los Angeles, and Houston between comparable February and April observations during the outbreak.The selected snapshots use matching day-of-week and time-of-day conditions with clear skies.
- Supplementary Note SN-2: Description of the Mobile Device Location Dataset: Mobile-location data provide social-distancing metrics and point-of-interest visit patterns, aggregated by county to characterize stay-at-home behavior, work-site activity, and retail mobility.The source aggregates anonymized GPS data from numerous applications, while retail mobility sums daily visits to selected retail POIs.
Supplementary Note SN-4: Data Sources for the COVID-EMDA+ Data Hub
The COVID-EMDA+ data hub integrates electricity-market, weather, satellite-imagery, public-health, and mobile-device location data to characterize factors influencing electricity consumption during COVID-19. Confirmed case counts are used to represent pandemic severity because of their availability, popularity, and modeling performance.
- Data sources: Electricity-market data are sourced from CAISO, MISO, ISO-NE, NYISO, PJM, SPP, and ERCOT, with EIA and EnergyOnline data used to improve quality and fill missing values.The data include load, generation mix, and day-ahead locational marginal price.
- Data sources: The hub uses Iowa State University weather observations, NASA’s daily 500-meter VNP46A1 Black-Marble imagery, Johns Hopkins county-level COVID-19 counts, and SafeGraph anonymized GPS data.Weather variables include temperature, relative humidity, wind speed, and dew temperature; SafeGraph provides social-distancing and points-of-interest visit datasets.
- Data sources: The hub combines electricity-market, weather, satellite-imagery, COVID-19 public-health, and mobile-device location data to capture multidimensional influences on electricity consumption.Electricity data cover load, generation mix, and day-ahead LMP across seven U.S. wholesale markets, supplemented by EIA and EnergyOnline data.
- COVID-19 severity metric: Confirmed COVID-19 cases are selected over deaths, hospitalizations, and ICU occupancy because they offer higher availability, popularity, and modeling performance.The paper notes that substituting hospitalizations or deaths does not change the key findings, while confirmed cases are easier to collect and deaths are delayed indicators.
Supplementary Methods · Supplementary Method SM-1: Pre-estimation Preparation · Supplementary Method SM-2: Restricted VAR Model Estimation
The supplementary methods prepare electricity, health, weather, and mobility variables for restricted VAR estimation by transforming and statistically validating candidate time series. They then estimate load-reduction dynamics with lagged variables and impose Granger-based constraints to exclude undesirable causal relationships.
- Supplementary Methods: The restricted VAR estimation and associated statistical tests and analyses are implemented with the Statsmodels module in Python.The implementation follows the described coefficient-estimation and model-verification steps.
- Data Pre-Processing: Candidate inputs combine electricity load reduction with COVID-19 cases, stay-at-home population, home dwell time, on-site workers, and retail mobility.Load reduction samples with negative reduction are dropped; listed variables are logarithmically transformed where specified.
- Augmented Dicket-Fuller Test: Each candidate time series is tested with the Augmented Dickey-Fuller test and differenced to improve stationarity before VAR calibration.The ADF test assesses whether each series is non-stationary and possesses a unit root.
- Cointegration Test: Cointegration testing removes selected input tuples when original non-stationary series exhibit long-term correlation, making them unsuitable for restricted VAR modeling.The test complements ADF testing by examining long-term relationships among the original inputs.
- Granger Causality Wald Test: Granger causality tests assess whether lagged values of one time series probabilistically affect the current value of another.The test excludes concurrent and future values as causes of the current target value.
- Supplementary Method SM-2: Restricted VAR Model Estimation: After statistical verification, the method models load reduction with a p-order Vector Autoregression using de-trended stationary inputs without long-term correlation.Load reduction is the target output, while selected parameter variables include COVID-19 cases, completely stay-at-home rate, and median home dwell time rate.
- Supplementary Method SM-2: Restricted VAR Model Estimation: The restricted VAR coefficients are estimated separately for each variable with Ordinary Least Squares and aggregated into regression matrices across lags.The matrices are formed from the estimated coefficients A_k for 0 ≤ k ≤ p.
- Supplementary Method SM-2: Restricted VAR Model Estimation: Granger-test evidence can impose OLS constraints that eliminate undesirable causal relationships from the restricted VAR model.The example restricts a relationship in which load-reduction variation would cause changes in stay-at-home population.
Supplementary Method SM-3: Restricted VAR Model Verification · Supplementary Method SM-4: Post-estimation Analysis
The restricted VAR model is verified through residual stationarity, autocorrelation, stability, and robustness checks before post-estimation interpretation. Post-estimation analysis uses impulse responses and forecast error variance decomposition to characterize dynamic responses and variance contributions.
- Supplementary Method SM-3: Restricted VAR Model Verification: The restricted VAR verification requires stationarity, non-autocorrelation, and normality of residual data, with residuals defined from the model’s observations and estimated coefficient matrices.The residual equation uses the observed vectors, residuals, maximum lag, and estimated coefficient matrices.
- Augmented Dicket-Fuller Test for Residual Stationarity: The ADF test checks whether residual data are non-stationary and possess a unit root.
- Ljung-Box Test for Residual Autocorrelation: The Ljung-Box and Durbin-Watson tests assess residual autocorrelation, including group serial correlation and lag-1 dependence.The Ljung-Box test uses a 5% significance level with 40 selected lags, while Durbin-Watson tests first-order autoregressive dependence.
- Stability Test: The stability test classifies a restricted VAR model as stable when the absolute values of all eigenvalues are equal to or less than 1.The test is described as a unit root test based on the model’s eigenvalues.
- Robustness Test for Parameter Stability: Robustness is evaluated by perturbing input parameters and applying statistical tests such as AIC, selecting a model that balances validation accuracy with robustness.
- Supplementary Method SM-4: Post-estimation Analysis: After the model passes statistical tests, post-estimation analysis interprets its results through impulse response analysis and forecast error variance decomposition.Impulse responses describe variable evolution after shocks, while FEVD attributes forecast-error variance to shocks in other variables.
- Impulse Response Analysis: Impulse response functions forecast the dynamic behavior of electricity consumption after changes in exogenous factors and can combine with public-health mobility models to simulate policy impacts.The impulse response is initialized with a user-defined unit element in R(0).
- Forecast Error Variance Decomposition: FEVD quantifies each variable’s forecast-error variance share attributable to shocks in other variables using the VAR companion form, forecast errors, and Cholesky factorization.The decomposition defines w_ij(h) as the proportion of the i-th variable’s h-step forecast-error variance accounted for by shocks to the j-th variable.
Supplementary Method SM-5: Restricted VAR Model Selection
The method uses a city-specific numerical search over Restricted VAR inputs, training dates, model order, and coefficient-zeroing rules. It evaluates each parameter combination and selects the optimum using AIC, BIC, and explained variance in load reduction.
- Restricted VAR Model Selection: City-specific numerical search selects Restricted VAR inputs, training dates, model order p from 1 to 7, and coefficient-zeroing rule r.The search addresses the absence of an explicit rule for selecting parameter variables and the training-data range.
- Restricted VAR Model Selection: Three coefficient-zeroing rules are considered, including Granger Causality Wald Test thresholds of p > 0.1 or p > 0.05.All rules first set to zero the relevant coefficients in the first column of each matrix, except the first row.
- Restricted VAR Model Selection: Each candidate combination runs the preceding procedures, including differencing, stationarity and cointegration checks, Granger causality testing, Restricted VAR estimation, and residual diagnostics.Model performance is quantified with AIC and BIC information criteria.
- Restricted VAR Model Selection: The optimal combination minimizes AIC and BIC while explaining a large proportion of variance in load reduction.The selected parameters finalize the model after the iterative search.
Supplementary Tables
The supplementary tables document restricted VAR model settings, diagnostic tests, COVID-19 metric availability and correlations, and comparative statistical results across COVID-19 indicators.
- Supplementary Tables: Table 1 lists each city’s restricted VAR training dates, lag orders, and hyperparameters.The table also references the Rule 2 definition in Supplementary Method SM-5.
- Supplementary Tables: Table 2 presents city-level restricted VAR diagnostic tests, including ADF, cointegration, Ljung-Box, Durbin-Watson, and stability tests.It reports p values for ADF and Ljung-Box tests, Durbin-Watson statistics, and Boolean outcomes for cointegration and stability.
- Supplementary Tables: Table 3 summarizes the availability of COVID-19 metrics in hotspot cities and their correlations with confirmed case data.The supplementary implementation distinguishes metrics based on confirmed cases, deaths, hospitalizations, and ICU occupancy.
- Supplementary Tables: Table 4 reports restricted VAR results for different COVID-19 indicators using AIC, BIC, explainable rate, and impulse-response measures.Lower AIC and BIC indicate better model performance; the explainable rate measures variance attributable to factors other than electricity consumption’s own trend.