Once a minute, the frequency on a power grid dips by a few thousandths of a hertz. Then it recovers, and 60 seconds later it dips again.
The dip appears in grid data from Europe, Australia, the United States and Malaysia. On systems thousands of kilometres apart, it arrives at the same instant.
Tennet Germany, the country’s largest transmission system operator, brought the anomaly to Dirk Witthaut at Forschungszentrum Jülich. His team spent years on it. What they published in Nature Energy is a piece of battery software behaviour, running on a clock the whole world shares.
Frequency is how a grid reports that generation and demand have parted. Every synchronous machine on a network turns at a rate tied to that number, and the number has to stay inside a band that operators set with some margin, because the cost of leaving it is not linear. Small deviations are routine and absorbed by reserve. Larger ones wear equipment down. Oscillations that an operator cannot explain, or answers with the wrong corrective action, can end in a blackout, which is why an anomaly nobody could account for drew attention disproportionate to its size.
Hydropower turbines take the damage first. Every turbine vibrates, and vibration at a resonant frequency amplifies instead of fading, so a plant that has run for decades can develop a mode nobody designed for. Witthaut’s group started there. They modelled a hydropower control system on the Scandinavian grid and found settings that produced a resonance peaking once a minute. The model could generate the signal. It could not explain why the signal never stopped.
Mechanical resonances come and go.
Anything that regular needs a clock, the team reasoned, because nothing in a turbine spins to a one-minute beat by accident.
They installed more than 80 power quality meters around the Jülich campus, a site with medium- and low-voltage lines, transformers, solar panels, battery storage wired in as uninterruptible supply, a high-performance computer and the usual office machines on smaller backup units. The meters sampled frequency, voltage, current and power every three seconds, which is fast enough to catch a one-minute period without recording the whole waveform, and cheap enough to deploy across a site rather than at a single point.
The transformer attached to the battery system peaked every minute.
Stationary batteries poll their state of charge on a one-minute cycle, from the cabinets behind a data hall to grid-scale installations measured in megawatt hours. A single polling event is invisible on a national grid. These devices are internet-connected, though, and their clocks are close enough to identical that thousands of them poll inside the same second.
Why one minute, and not five or fifteen, is not something the published account explains. A shorter interval buys finer visibility of remaining runtime; a longer one costs less in network traffic. Whatever the reason each vendor settled on, the intervals converged, and convergence is the whole problem.
Ireland supplied the test case. Data centres reached 18% of Irish grid load by 2022, and each one runs a fleet of uninterruptible supplies. Between 2016 and 2021 the minute-by-minute dip on the Irish system grew from 3 millihertz to 12 millihertz, a fourfold increase in five years. Some recent readings pass 20 millihertz.
The Irish grid runs at 50 hertz and is built for swings of plus or minus 200 millihertz. The operator would rather hold inside 100. Twenty is a small share of that headroom, and headroom is shared with every other load that pushes frequency around.
That series is the strongest part of the case, because the two variables move together. Data centre capacity on the Irish system grew across the same five years, and the dip grew with it, from 3 millihertz to 12 and then past 20 in some readings. Correlation of that kind is not proof on its own. It becomes persuasive when read next to the campus measurement, where the suspected source sits on one transformer and the meters can watch it directly.
What makes the finding usable is the instrumentation, not the headline number. Eighty-odd meters sampling every three seconds is a local measurement, and the researchers had to show it matched the behaviour of national grids. The matching peak on the campus transformer did that, and the Irish series, drawn from operator data rather than a model, carried the argument to transmission scale.
The remedy is cheap to describe. Randomise the polling schedule so that batteries stop checking on the same second. Witthaut’s team recommends that makers of uninterruptible supplies and grid-scale batteries change the software that sets the cycle.
Randomisation is also the kind of fix that looks trivial and meets procurement. An uninterruptible supply is bought for runtime and service terms, and in most deployments the firmware that sets the polling interval belongs to the manufacturer, not to the operator running the hall. A data centre owner who wanted the change would have to ask for it, which is why the recommendation lands on vendors and not on their customers.
Other equipment carries the same signature. The researchers tested a smart oven and clusters of Raspberry Pi computers and found periodic power fluctuations in both. That fix is also a software change, and much of that software goes unpatched: research cited in the paper found fewer than 3% of internet-connected devices running the latest firmware.
Regulators can reach the equipment connected to their own networks. The pattern crosses borders, so a national rule leaves the rest of it standing.
Two things stay open. Battery makers have not said whether they will randomise their polling, and no published figure yet shows how much of the dip would disappear if every one of them did. The second is measurement: a grid operator can watch its own frequency trace before and after a supplier ships firmware, and that comparison, repeated across a few markets, would settle whether the mechanism holds at scale.
Until then the minute hand keeps its grip. The dip is small enough to ignore on most days and regular enough to be found in any frequency log an operator cares to search.