A grid tells you when society changed

Focused brief on: “A global hourly electricity demand dataset for studying power system resiliency during COVID-19,” Scientific Data, published 3 August 2026.

The plain-English version

A power grid has to make electricity at the same moment people use it. That real-time need is called electricity load.

Most people talk about demand as a total: how much electricity a country used in a month or a year. This dataset keeps the clock. It asks: when in the day did demand rise, fall, flatten, or move after COVID lockdowns changed where people worked and when buildings were active?

That matters because a grid can survive a lower total demand and still be stressed if the timing is wrong. Batteries, backup generators, demand-response programs, and grid operators care about the shape of the day: morning peaks, evening peaks, and low-use hours.

The useful thing here is not “COVID changed electricity use.” That was already known. The useful thing is that someone packaged the change into a reusable, cross-country dataset.

The read

This is measurement infrastructure. The paper turns the COVID lockdown period into a comparable stress test for power systems across 35 countries.

The core object is a typical daily load profile: a 24-number shape that says how much electricity demand normally occurs at each hour of the day for a given country, season, day type, and period.

The authors split the world into three time regimes:

Then they ask how the daily load shape changed between those regimes.

My read: the dataset is more interesting than the paper’s framing. The paper frames it as COVID energy-resilience data. I would frame it as a reusable template for turning a social shock into a grid-shape dataset.

Terms worth defining

Electricity demand / load: the amount of electrical power being used at a specific moment. If everyone turns on air conditioners at 5 p.m., load rises.

Load profile: the shape of demand across time. A daily load profile has 24 hourly values. It tells you not only how much electricity was used, but when.

Peak: the highest demand point in the day. Peaks are expensive and operationally important because the grid must have enough generation and delivery capacity ready at that moment.

Trough / off-peak / base load: the low point in the day. This matters because grids still need stable operation when demand is unusually low.

Demand response: programs that shift or reduce electricity use at useful times, for example by delaying industrial loads or adjusting thermostats.

Medoid: the real observed day that best represents a group of days. It is like an average, but it chooses an actual day instead of creating a synthetic smoothed curve.

Stringency Index: the Oxford COVID-19 Government Response Tracker score that tries to quantify how strict a government’s COVID rules were on a given day.

What the paper and dataset actually contain

Source paper: “A global hourly electricity demand dataset for studying power system resiliency during COVID-19.”

The paper says the dataset collates and harmonizes hourly electricity demand for 35 countries, with data windows ranging from as early as 2015 through 2024, depending on country/source availability.

The actual typical_daily_usage_patterns.csv I inspected contains 978 typical-profile rows. Each row has a country, season, period type, day type, and a 24-entry daily profile.

Countries represented in the typical-profile file:

Austria, Australia, Belgium, Bulgaria, Brazil, Chile, Cyprus, Czech Republic, Germany, Denmark, Estonia, Spain, Finland, France, United Kingdom, Greece, Croatia, Hungary, Ireland, India, Italy, Japan, South Korea, Lithuania, Luxembourg, Latvia, Netherlands, Poland, Portugal, Romania, Sweden, Slovenia, Slovakia, United States, and South Africa.

How they built it

The pipeline has four main steps.

First, the authors gather policy-stringency data from OxCGRT for 2020 and 2021. That gives them a daily government-restriction score for each country.

Second, they collect hourly electricity demand. For Europe except the UK, they use ENTSO-E. For India, they use the India Climate & Energy Dashboard. For the UK, Australia, the United States, Japan, South Africa, South Korea, Brazil, and Chile, they use the International Energy Agency dashboard / API. The paper says they rejected an initial UK ENTSO-E source after visual checks showed post-lockdown figures that looked underreported and skewed, then used IEA data instead.

Third, they define the strictest lockdown period country by country. The rule is relative, not global. For each country, they take the dates where one of the country’s top two stringency-index values was in effect. The start is the earliest such date; the end is the latest such date. This avoids saying every country had to cross the same absolute policy threshold.

Fourth, they compute representative daily load profiles. They segment by:

For each segment, they choose a medoid daily profile. That means they pick the actual day whose 24-hour curve is closest to all the other days in that segment.

This is a practical choice. It preserves a real day’s sequence instead of smoothing away the hour-to-hour shape.

What the schema looks like

The Zenodo record includes PostgreSQL dumps, CSV files, and processing scripts.

The important tables/files are:

The key row in the final analytical table is conceptually:

“For Austria, on summer weekends, in the post-COVID period, here is the typical 24-hour electricity-demand curve.”

That is much more reusable than one national time series or one COVID case study.

What the validation says

The paper uses independent two-sample t-tests to check whether the load-profile metrics differ across periods.

The metrics are not only total load. They focus on shape:

The paper reports statistically significant changes across many countries and periods. Examples in the manuscript include:

That heterogeneity is the point. Lockdown did not produce one universal electricity signature. The same social shock hit different economies, climates, housing patterns, and industrial mixes differently.

Why this matters

The strongest insight is simple: resilience is a timing problem before it is a totals problem.

A country can use less electricity overall and still create a harder grid-management problem if demand moves into inconvenient hours. A battery fleet, for example, is not sized only by annual energy. It is sized by when it must charge and discharge. A demand-response program is not useful because people use “less electricity” in the abstract. It is useful because it moves demand away from the wrong hour.

COVID lockdowns created a rare natural experiment. Factories slowed. Offices emptied. Homes became daytime workplaces. Schools closed. Mobility changed. Those changes pushed electricity demand around inside the day.

That makes the dataset useful for questions like:

The non-obvious use

The obvious use is energy research.

The better use is stress-test design.

This dataset gives a model for how to package a shock:

  1. define the shock with an external index;
  2. align operational telemetry to the shock window;
  3. preserve time-of-day shape rather than only totals;
  4. normalize enough for cross-country comparison;
  5. leave hooks for weather, GDP, public-health, and policy data.

That pattern transfers beyond electricity.

For a software or agent system, the equivalent would be: take an external disruption, align runtime telemetry to it, preserve the shape of behavior over time, and compare before/during/after. The principle is the same. The timing of behavior often carries the signal that totals hide.

What I would not overclaim

This is not causal proof that lockdown policy itself caused every load change.

The authors use a standardized lockdown definition and t-tests. That is useful, but it does not fully control for:

The medoid method also trades causal precision for comparability. It gives a transparent representative day, not a full counterfactual model of what demand would have been without COVID.

The dataset is therefore best treated as a clean baseline and reusable substrate, not as the final causal analysis.

What I would do next with it

If I were using this seriously, I would run three follow-ups.

First: plot the daily profiles for a few countries across pre / lockdown / post periods. Start with India, South Korea, Australia, Luxembourg, and the United States. The paper already flags those as showing different kinds of behavior.

Second: join the profiles to temperature data. That would separate behavior change from weather-driven load. This matters because heating and cooling can dominate demand shape.

Third: turn the shape changes into operational metrics: peak reduction, peak shift in hours, trough change, morning-evening spread, and ramp steepness. Those are closer to what grid planning and battery sizing actually need.

Bottom line

This is a good dataset because it keeps the clock.

The COVID shock did not just change how much electricity countries used. It changed when they used it. That timing layer is where grid resilience lives.

The caveat is that the dataset is an organized signal, not the complete explanation. Weather, sector mix, and policy enforcement still need to be modeled. But as a reusable substrate for asking sharper questions, it is genuinely useful.

Sources