Whose measurements these are
The National Energy System Operator publishes the pulse of a country’s electricity, once a second, for anyone to use.
NESO, which before 2024 was National Grid ESO, publishes the system’s frequency every second and its inertia every half hour under an open licence, and Elexon publishes what each kind of generation and each power link was carrying, every five minutes. Read together, they show how much power was lost in each trip and how the whole system answered.
- Who publishes them
- The National Energy System Operator and Elexon
- The data
- From NESO (before 2024, National Grid ESO): the system frequency, one value a second, and the outturn inertia, one value each half hour. From Elexon: generation by fuel type every five minutes, BMRS dataset FUELINST.
- The years
- Frequency from April 2017 to September 2026, in which forty-one trips were found between July 2017 and May 2026.
- Where they are published
- NESO’s data portal, www.neso.energy/data-portal, and Elexon’s BMRS, bmrs.elexon.co.uk.
- The acknowledgments they ask for
- Supported by National Energy SO Open Data. Contains BMRS data © Elexon Limited copyright and database right 2026.
What happened
When a major power link suddenly drops, the electricity of a whole country stumbles for about a minute.
Forty-one times in nine years, a link bringing power into Great Britain cut out, and the system’s frequency fell and came back.
Every generator in England, Scotland and Wales turns in step, fifty times a second. What holds that number steady? Supply and demand, matched moment by moment. Take away a thousand megawatts in an instant and the whole system slows.
That is what happens when an interconnector trips. Within seconds the frequency sags, power held in reserve for the purpose pushes back, and inside a minute the fall has been caught.
The operator’s once-a-second record holds every one of these stumbles. A written rule found them in 114 months of it, and Elexon’s figures gave the size of each loss, from 693 to 1,402 megawatts.
What the model did
Twelve real stumbles it had never seen, and every mark met.
The model treats the whole island as one spinning mass, with one reserve that answers a falling frequency. Its three fitted constants came from twenty-eight of the trips. Every third trip in time order was held back, and the passing grade was fixed before any model was run.
On the held-back trips the model said how far the frequency would fall, how steeply, and where it would sit a minute later. Twelve were scored, and all seven marks were met. One was set aside before scoring, by a rule written beforehand, because two links had failed at once.
How close was it? On the depth of the fall its typical error was about 17 percent. The model’s description is plain about what that is worth: a one-line rule, depth in proportion to the loss, scores about 18 on the same twelve, and those twelve were an easier set than the rest.
Met
Where it stops
It carries one way of answering across nine years of a system that kept changing. Outside the losses and frequencies it was compared on, it stops and says why.
What it was tested for
- Great Britain as one frequency, for the sixty seconds after one importing link trips.
- Losses of 693 to 1,402 megawatts.
- Frequencies no lower than 49.55 hertz.
Past that, and what its own description says of itself
- A loss outside that range, a frequency under 49.55 hertz, or a second loss during a run: the model stops and names the reason.
- Its errors follow the calendar. Before 2021 its falls are about a quarter too shallow, after 2023 too deep, and the six trips since November 2025 run 39 to 80 percent too deep.
- The system changed in those years, with new response services from 2020 and a lower minimum inertia in 2024, and the model has one gain and one lag for all of it.
- The response the operator held on each day is not in the open data, so it is not an input.
- On the trips it was fitted on, with each left out in turn, it missed its mark on depth.
- There is no network in it, no generator that trips as a consequence, and nothing after sixty seconds.
For the reader who wants every number
In the verifier’s own words.
Everything above is this page’s telling. What follows is the verifier’s own, exact and whole: four parts of the description of the model energy/grid_gb, as read on the morning of 11 October 2026. Where the verifier says card it means a model, where it says line it means a pass mark, and where it says refuses it means the model stops there and says why.
The real thingobject
The Great Britain synchronous area, the 50 Hz transmission system of England, Scotland and Wales, in the 60 seconds after an importing interconnector, or one pole of it, trips: 41 trips of 693 to 1402 MW between July 2017 and May 2026.
The passing grade, fixed firstpass_line
Fixed before any model was run (protocol SHA-256 35c91fa9fe440ad852390b77bcd6ecddd8e7029630e620a660b11aad1f481ea9, amended once before the structure of this card was fitted, SHA-256 0fd449fc45839ef37928db97995b43f47cb22bb2ae89d531ccf4592e6e9732ba). Depth of the fall and the steepest second: root mean square of the relative error at most 0.25 and mean within 0.10, and the depth error smaller than that of the rule that depth is proportional to loss. Frequency near 60 s: root mean square error at most 0.06 Hz. The whole trace: median root mean square error at most 0.05 Hz.
The resultresult
Compared with the held-back trips on 2026-10-09, run once through the engine: all seven lines were met on 12 trips. One of the 13 was dropped before scoring by the quality rule written beforehand, the trip of 2 September 2021, whose recorded frequency moves 0.275 Hz in the ten seconds before the fall. That was two opposite losses at once, an export link and an import link, and not one trip. Depth of the fall: root mean square relative error 0.166 and mean -0.018, against 0.182 for the rule that depth is proportional to loss on the same trips; the worst was 37 percent too shallow, on 9 May 2019. Steepest second: 0.103 and mean +0.043. Frequency near 60 s: 0.046 Hz. The whole trace: median 0.036 Hz. No trip was refused by the card. The held-back twelve were the easiest third of three: on the other two thirds, each with constants refitted on the rest, the depth error is 0.258 and 0.326 with three lines missed each, and 18.7 percent of random twelves of the forty meet all seven lines. The card's margin over one-constant rules on the twelve is 0.015: depth proportional to loss 0.182, a fall to one fixed frequency 0.180, a straight line 0.171, the card 0.166. The loss and the lowest frequency of every trip, the held-back ones included, were on screen before the protocol was fingerprinted. A further 39 trips of the same kind, which the scan's rule had dropped because it reads the fall at the first second to reach 0.12 Hz, were found by the second scorer; 36 were inside the card as served through engine 0.6.0 and were scored after the fact, and six of them, of 500 to 692 MW, are under the losses the card takes from engine 0.7.0. All 36: depth 0.269 and near 60 s 0.067 Hz, two lines missed. May 2018 to February 2025, 23 trips: depth 0.136, all lines met. 2024 to 2026, 15 trips: depth 0.345, six lines missed. The six trips since November 2025 run 39 to 80 percent too deep. The binary that made the scored runs (SHA-256 da8ee131279452aa4f1f56afb71e801f622a050882554fb6d632dd23f9b7384c) was overwritten by a later build; the later build returns all 41 saved answers byte for byte. A second scoring of 2026-10-09, on the same machine and from the scorer's own scan of the monthly files, found the same numbers: evidence/checks/second_party/V-32/.
Its limitslimits
The held-back trips are now spent, and they are an easier set than the opened ones: the proportional rule scores 0.18 on them and 0.337 on the opened 28 with the card's constants (0.35 with each trip left out of the fit in turn), which hold a trip followed by a second loss (22 December 2023) and the shallowest fall of all (9 May 2026), so the card's margin over that rule on the held-back trips is small. On the opened trips the card missed its mark: with each trip left out of the fit in turn the depth error was 0.302 against 0.25 and the error near 60 s was 0.097 Hz against 0.06, while the steepest second, 0.154, and the trace, 0.050 Hz, met their lines and the depth beat the proportional rule, 0.350. No structure of the protocol met every line; this is the one with the smallest depth error. The response the operator held on each day is not in the open data and is not an input: the card carries one response for nine years. Its errors follow the calendar: before 2021 its falls are about a quarter too shallow, after 2023 too deep. The depth error moves by +0.066 a year (standard error 0.013), from -0.30 in 2018 to +0.17 in 2025, nearly twice the 0.25 mark across the span. The system changed: Dynamic Containment from October 2020, 500 MW rising to 1,576 MW by July 2023; Dynamic Moderation and Dynamic Regulation from March 2022; the minimum inertia lowered in 2024. The card has one gain and one lag for all of it. The inertia is the operator's estimate, which leaves out the Stability Pathfinder units. It is the sharper input: 0.9 or 1.2 times the published inertia misses the line on the mean of the steepest second. The size of the loss is steady to 0.35 percent by the steadiness rule. The range was narrowed on 2026-10-10, after the scoring and with every trip seen, so the narrowing is not a new test. Losses: every scored trip lost 693 MW or more. The six trips of 500 to 692 MW inside the other spans, among those the scan lost, run through the engine by the protocol's lines, missed three of the seven: depth 0.2512 against 0.25, and against 0.2509 for the proportional rule, steepest second 0.273 against 0.25; the mean depth error -0.032, the steepest second's mean +0.057, near 60 s 0.054 Hz and the trace median 0.030 Hz met theirs. A loss under 693 MW is refused as LossOutsideCard. Frequency: no compared trip went under 49.553 Hz, so a frequency under 49.55 Hz is refused as FrequencyBelowCompared, where through engine 0.6.0 the card went on to 49.46 Hz with a flag. Two numbers are not fitted. They were first written from memory, and the second scoring found their sources. Demand's relief of approximately 2.5 percent per hertz: National Grid, System Operability Framework 2016, p. 70, and the NESO webinar of 12 March 2025; the documented range is 2.0 to 2.5. The 49.8 Hz knee is Dynamic Containment's (NESO, Response Services Service Terms v6; Dynamic Containment testing guidelines, August 2020): right for that one service from October 2020, and before it there were static blocks at 49.7 and 49.6 Hz. With the three constants refitted, a relief from 0 to 5 percent per hertz changes no verdict, and the pass holds for a knee of 49.80 to 49.90 Hz and is lost at 49.75. A run is one loss: a loss_mw in a later segment that is not the first segment's is refused as LossChangedDuringRun, where through engine 0.5.0 it was taken as the total level of loss from then on. No changing loss was compared with a measurement: the one measured trip with a second loss, that of 22 December 2023, is the card's worst. The recomputation outside the engine failed against its criterion as first written, in 39 of 80 runs: only the response and the two integrals were outside, the response by at most 4.2e-10 MW, and the frequency was never above 0.05 of the bound. The yardstick was changed afterwards by the card's author. A run of exactly 60 s in decimal lands every step: 3,000 steps of 0.02 s land 3,000 and read 60.00000000000378 s. Through engine 0.5.0 the last was refused at some step sizes, because the clock is summed in binary; the comparison with 60 s now carries the published rounding allowance, and no answer that landed before has changed. Known defect of the scan that found the trips: its rule of 0.15 Hz or more lost four falls of exactly 0.150 Hz to rounding. One frequency for the whole area; no network, no generator that trips as a consequence, nothing after 60 s. One party wrote the card and ran it. It was scored twice, the second time by a second scorer, on one machine.
Hand this one to your agent.
https://mcp.zerotrustphysics.com/mcp
Something to paste
Connect to https://
Give it this page too.
What it will find
This model’s own description: what it takes in, where it stops, the trips it was held to and how it did, in its own words.
Give the address and this page to your own agent first. Then write or call.
- Write
- [email protected]
- Call
- 928-264-7883