Power grid · Frequency

The power grid reports its own blackouts — and the compressor catches them for free

Grid frequency wobbles slightly around its nominal value all the time. When a plant trips, it jolts — and "how much the compressor has to speak" gives the event away, with no detector programmed.

The demo in one sentence We took 6 real months of the British system frequency (measured every second) and ran it through TUBE, our compressor with a guaranteed error bound. The grid stays in a narrow band around 50 Hz, so the compressor is quiet 94.7% of the time. On every frequency excursion — a load/generation imbalance — it starts to "speak", and the emission rate jumps. Event detection comes out as a byproduct of compression, and certified, because the norm that defines the band is the tolerated error itself.

1The scenario

Grid frequency — 50 Hz in Europe, 60 Hz in Brazil — isn't fixed: it wobbles slightly all the time, following the balance between what's consumed and what's generated. If load suddenly rises (or a plant drops out), frequency falls; if there's surplus generation, it rises. That's why the system operator chases this balance second by second. And there's a norm defining the acceptable band (e.g., staying within ±0.2 Hz of nominal in normal operation) — so the tolerated error is already written in a document.

2The "tube" and the "deadband"

TUBE compresses data with a simple rule: instead of storing every measurement, it draws a "tube" of tolerance around the signal and stays quiet as long as the value doesn't leave the tube — whoever reconstructs it later knows it was around there, within the guaranteed margin. That silence is the deadband. Only when the signal breaks out of the tube does it emit a new point.

In a healthy grid, frequency drifts slowly inside the norm's band — so the compressor barely speaks. It's the perfect deadband case: 94.7% silence across six months.

3The demo

We ran the 6 months of 2026 (January to June) of the British system frequency, every second — 15.6 million measurements. While the grid stays in band, the tube holds. But on every excursion (a generator trip, a load step) the frequency jolts out of band and the tube breaks — the emission rate jumps. In the chart below, each peak is a real event of the semester; the largest took the frequency to 49.6 Hz.

100%0%janfevmarabrmaijun excursão 49,6 Hz
Figure 1. The compressor's "voice" over six months: near-silence, with a clean peak at every grid frequency event. Open the interactive demo, with the real data and the frequency dip.
Measured on the codec: deadband over the semester 94.7%; in the event window 88× vs. raw with the maximum error within the bound; the surprise jumps 14× above rest.  ▶ open the interactive demo

4Why this is beautiful

Nobody programmed a grid-event detector. The compressor was just trying to save space. But "how much it needs to speak" is, in practice, a measure of surprise: grid in balance = silence; frequency jolt = chatter. So event detection comes out as a byproduct of compression — with a rare advantage: because the norm defines the band (the ε), the alarm is certified, not a guess. It's the same argument as the submarine cable that became a seismograph, now on the power grid — and the direct sibling of our synchrophasor (PMU) note, where the phasor error (TVE) is already the norm's ε.

Honesty

We measured on the British system (50 Hz), because it's the cleanest open 1-second public series. In Brazil the grid is 60 Hz and the behavior is the same — the physics of the argument doesn't depend on the nominal value; what matters is "signal in band + event = surprise". High-resolution raw 60 Hz data is still barely open here. And unlike partial discharge (a noisy signal where the raw surprise doesn't isolate the event), frequency is the clean case: slow signal, norm band, crisp event.