Skip to content

Experiment 194: D8's drops follow one engine whose output stopped being added back

We asked whether the long stretches in which D8 (one of the 16 Primary substations in the trial) draws less power than a fitted model expects come from one 6 MW gas engine, labelled G2 on this page. NGED's notes say that G2's output is added to D8's reading, so that D8 shows demand plus the engine's output, and that G2's own telemetry channel (its analogue) does not work. We reran the design note's calculations from committed code. We call the share of half-hours in which D8's metered power sat more than 2.5 MW below the fitted model's expected power the "shortfall share". We found the date on which that share jumped. We also checked the same share on three comparison Primaries (D3, D5, and D14) with no known engine.

The design note's numbers reproduce, and its account of D8's drops holds in outline. While G2's output arrived, D8's reading rose with G2's output (D8's reading rose by 0.02 MW for each 1 MW of G2 output; had the output not been added back, D8's reading would fall by 1 MW for each 1 MW). The shortfall share stepped from 0.010 to 0.257 in the week of 2023-08-14, the week G2's last output reading arrived. It stayed between 0.10 and 0.46 in every month to January 2026. The shortfall share of the three comparison Primaries stayed near zero.

Two findings differ from the design note's wording. G2's readings did not stop arriving: 98 later days carry all-zero readings, and D8's shortfall share is the same with a zero reading present as with none. D8's shortfall also fades from March 2026 with no matching change in G2's readings (they already stopped on 2026-01-26). Experiment 191's scored year (2025-07-01 to 2026-06-30) therefore holds both high-shortfall and low-shortfall months.

Figures

Weekly shortfall share of D8 and three control substations

Figure 1. D8's weekly shortfall share against D3, D5, and D14 (near zero throughout).

Day-by-day shortfall share, G2 daily maximum and G2 reading count

Figure 2. The break by day: D8's shortfall starts on the day G2's last output reading arrives.

G2 running share by price fifth

Figure 3. G2 ran in 9% of the cheapest fifth of half-hours and 26% of the dearest.

D8 residual histograms before, during, and after G2's output weeks

Figure 4. D8's residual in four periods. The negative tail after August 2023 is broad, not clearly quantised in units of 2 MW.

D8 monthly shortfall share with experiment 191's scored year

Figure 5. Monthly shortfall share. The blue band is the scored year of experiment 191.

Reproduction of the design's numbers (PS1)

  • Stop rule: passed. 2,054 half-hours join to the system price, exactly the design's count.
  • G2's output (2023-07-03 to 2023-08-15, 44 days): units running 0, 1, 2, 3 in 1,704, 137, 133, 80 half-hours (exact). Running share 0.170 (Wilson 0.155 to 0.187). Only 440 of G2's 441 positive readings fall in these weeks; the other is 0.2 MW on 2026-01-23.
  • Running by price fifth, cheapest to dearest: 0.089, 0.152, 0.151, 0.198, 0.259, against the design's 0.090, 0.151, 0.152, 0.191, 0.271 (the top fifth differs by 0.012). The cheapest and dearest fifths' intervals are 0.065 to 0.120 and 0.220 to 0.303. Spearman correlation of price with running 0.160.
  • By hour (UTC): 0.02 to 0.20 from 00:00 to 05:59 (design 0.02 to 0.10) and 0.25 to 0.28 from 19:00 to 21:59 (exact). The design's low range misses 05:00 UTC (0.20), where running starts.
  • Running spells (consecutive half-hours with G2 running): 56 (design 55), lengths 1 to 29 (exact) and median 4 half-hours (design 6; not reproduced, probably because a missing reading ends a spell here).
  • D8's residual while G2 ran, median at 1, 2, and 3 units running: over 2023-07-03 to 2023-08-15 we find +0.49, +0.43, and +0.23 MW. The design quotes +0.49, +0.53, and +0.23 MW for July, so the 2-unit value is not reproduced and the design's label ("July") does not match its first and last values. For July alone the medians are +0.19, +0.37, and +0.60 MW, which rise with output. All six medians are positive.
  • Shortfall share by month: May 2023 0.001 (design 0.000), July 0.000, September 0.353, October 0.357 (design 0.353), November 2025 0.147, January 2026 0.292, April 2026 0.008, May 0.022, June 0.039. The June and August 2023 values (0.035 and 0.112) are exact.
  • By calendar quarter, 2023Q4 to 2025Q4: 0.198 to 0.338 (design 0.197 to 0.343, which used quarters starting in September). 2023Q3 is 0.153 because July and August were mostly clean.
  • Springs: March to May 0.267 in 2024, 0.228 in 2025, 0.034 in 2026.

Prediction and decision rule

  • July 2023 at most 0.02: held (0.000, Wilson 0.000 to 0.003).
  • At least 0.10 in every month from September 2023 to January 2026: held, narrowly. The lowest month is February 2025 at 0.102.
  • Change point between 2023-08-10 and 2023-09-10: held. The weekly step is at 2023-08-14 (likelihood-ratio set and bootstrap 95% both 2023-08-14 to 2023-08-21).
  • No control within 30 days: held, with a caveat. D3 and D5 have no step (0.000). Their best split falls on the earliest allowed week, 2023-06-26, 49 days before D8's, because the likelihood is flat and the search returns the first tie. D14's step is +0.005 at 2025-09-29.
  • Two-step fit of D8 over the whole window: up at 2023-08-14 (0.010 to 0.254), down at 2026-03-09 (to 0.029).
  • Sensitivity: a threshold of 0.241 times each site's capacity equals -2.5 MW at D8, so D8's fit is unchanged. The controls stay below a step of 0.007, though D5's best split moves to 2025-11-10.

Does the account hold?

  • Gross while G2's output arrived: yes. Had D8 been net of G2, 10.2% of those half-hours would fall below -2.5 MW. The observed share is 0.0%.
  • Net afterwards: yes, from the shortfall's timing. The shortfall starts on 2023-08-15, when G2's readings stop mid-day (22 of 48 readings that day).
  • "Reading stopped arriving": only partly. G2 has 98 days with readings after 2023-08-24, all zero except the 0.2 MW. D8's shortfall share is 0.329 with a reading present and 0.270 with none (2023-08-25 to 2023-12-31; the intervals do not overlap), and 0.232 and 0.255 in 2024 to January
  • The difference has no consistent sign, and G2's later readings give no record of the engine running.
  • Quantised steps (output rising in units of 2 MW, one per running engine unit): not shown. The histogram's local peaks are at -4.1 MW and -3.1 MW, where steps of 2 MW would put them at -2, -4, and -6 MW. Noise in the expected power is wide enough to blur 2 MW steps.
  • 2026: the shortfall share falls from 0.292 in January to 0.070 in March and 0.008 in April while G2 has no readings after 2026-01-26. Either the engine stopped running, or NGED resumed adding the engine's output to D8's reading. The data cannot tell which.
  • Experiment 191's window (its constants: D8 data from 2025-04-01, scored runs 2025-07-01 to 2026-06-30): in those 12 months, 59% of half-hours fall before February 2026 (shortfall share 0.292) and 41% from February 2026 onwards (0.061). The window includes the 2026 change.
  • Register entries near G2 (MW and distance only): a 6 MW gas engine connected 2023-06-13 at 0.00 km in the ECR (under D14 in the register's Primary column); two Capacity Market gas-fuelled components at 0.05 km (six-figure grid reference, so about 0.07 km precision); a connected 10.98 MW battery at 0.27 km, registered under none of the 16 Primaries.

PS5: census

  • Series notes: 5 of 33 series carry an information note. Among the 16 Primaries, only D8 has an added-back note, and its note says that generators' analogues are not working. Of 13 metered generators and batteries, 4 carry a note: 1 solar analogue "not working", 1 solar analogue "suspect", G2 "not working", 1 biofuel customer "has not exported since June 2026".
  • Reading availability, 2024-01-01 to 2026-06-30: 11 of the 13 generators and batteries have readings in at least 74% of half-hours. G2 has 4% (last 2026-01-26), and 1 solar farm has none since 2024-09-26.
  • Shortfall share across all 16 Primaries: D8 is the only Primary with a sustained upward step of at least 0.10 (+0.246). D9's usable weeks start in January 2024 and open with a four-week burst (weekly shares 0.08, 0.52, 0.65, 0.34 from 2024-01-08), after which the share is near zero, so the fit reads a step of -0.198 at 2024-02-26. D18 has 27 usable weeks and one week at 0.301 in October 2025, fitted as a step of -0.038. D15 has no usable weeks. D14's highest week is 0.074.
  • ECR price-following assets under the 16 Primaries (connected): 2 gas or oil engines of 7.00 MW (6.00 MW at D14 from 2023-06-13, 1.00 MW at D5 from 2021-03-11) and 3 batteries of 0.65 MW (D11, D2, D9). 2 batteries (0.2 MW each, D16 and D3) are accepted to connect. 9 Primaries have none. The demand-side response register holds no rows. CHP: 2 slots, 1.40 MW, both at D10, none gas or oil engines.
  • Whole register: 168 connected gas or oil engines of at least 1 MW each, totalling 1,265.8 MW. Engines of this size are therefore not confined to the trial's Primaries.
  • PS5 prediction: held for sustained upward steps. D9's transient burst fits a step down of more than 0.10, and D15 and D18 lack the data for a fit, so "no other series" covers 13 of 16 Primaries fully.

Checks

  • Median minus mean forecast per site and mean-forecast error beside each contrast: not applicable, because the experiment scores no forecast.
  • Audit triggers: none fired. No control failed, the result matches its prediction, and no planned contrast exists to flip. Later experiments (PS2 to PS4) depend on this result, and this experiment is the replication the design asked for.
  • Caveats: the stored residual panel comes from experiment 130's frozen convex model driven by ERA5, so a changing model error shifts the shortfall share. The threshold of -2.5 MW is the design's; D8 is measured in apparent power (MVA, not MW), so a dip in apparent power at low demand can look like generation. The bootstrap interval resamples weeks within each side of the fitted split, so it is narrow by construction; the likelihood-ratio set is the more cautious interval.

What to try next

  • Ask NGED whether G2's output is still added back to D8, from when, and whether its analogue was repaired (design questions 1 to 3).
  • PS2 may use G2's 44 summer days as a known answer, and must treat D8 as gross before 2023-08-15 and net after.
  • PS3 and PS4 need an "engine visible in the target" flag, set to off from March 2026 until NGED answers.