Skip to content

Metered generators and anonymisation

NGED meters 6 solar farms, 3 wind farms, and 3 other generator or storage series in the trial area, and the output of every one of them may leave the private data store only under an anonymous label. This page lists the series, the export-cap record for the one generator under active network management, and what a plot or pull request may show. Public data about generators is treated differently, and the last section says how.

The series

  • The 32 series that carry data include 6 solar farms, 3 wind farms, and 3 other generators or storage units. The three are one battery energy storage system, one biofuel generator, and one series of type "Other (Generation)" (disk).
  • Of the 3 wind series, 1 is metered in MW and 2 in MVA (disk).
  • A type of "Other (Generation)" need not be solar. Do not assume a daylight profile (exp 182; one farm behind D8; provisional).
  • Read a farm's cleaned series from cleaned_solar_farms.parquet in data/ml-experiments/00179-*/ and keep rows with a null drop_reason. The file has no site column, so join studies.pv_dataset.pv_sites() on time_series_id for the label, the coordinates, and the capacity (exp 183; 6 farms; provisional).
  • Farms miss 7.6% to 27.6% of their half-hours. Check coverage before comparing farms (exp 179; 6 farms; provisional).
  • About 3.0% of the daylight half-hours of the six solar farms read exactly zero. Count a zero actual as inside a quantile forecast only by a rule stated in advance: the share at or below the 0.01 quantile (3.3% to 3.5% for research default 1) counts them, and the share strictly below it (1.2% to 1.4%) does not (exp 321; 6 farms; provisional).
  • A flat top in a farm's output is not proof of clipping. The cause is not settled (exp 179, 183, 218; 6 farms; provisional). See the metered solar farms theme.
  • A near-zero reading at a wind farm is not always low wind. Of the half-hours above 1.5 times cut-in wind speed, 1.6%, 13%, and 30% sit in the idle band at W1, W2, and W3, and ERA5 wind speed finds only part of them (exp 210; 3 wind farms; provisional).
  • UKV wind lowered the error of the three wind farms by about 0.2 points of capacity at 24 to 54 hours ahead, using a research archive and not the live feed, so the gain is an upper limit (exp 54; 3 wind farms; small gain; provisional). See the generator forecasting theme.

Active network management and the curtailment feed

  • One generator in the trial area is connected under active network management (ANM), and NGED has confirmed there are no others. A generator with no cap record therefore ran free (network page; docs).
  • Two records exist and they are different quantities. The setpoint history is one row per change of the export cap, in MW. The curtailment/ feed on NGED's S3 bucket is a derived volume of MW lost (data sources; docs).
  • The setpoint history is the one to use. The cap's sign, its start date, and the month the scheme goes live are on the network page (docs).
  • The export sits in data/studies/downloads/observations/NGED-ANM/, and studies.export_cap joins it onto an experiment's grid. It drops capped hours from training and keeps them at scoring (disk).
  • The clamp uses the cap in force, which a day-ahead forecast does not have, so it is an upper bound on what the cap could buy (code).
  • Curtailment at substations is scored by cost, not by this feed. The scoring code is a worked example on 2 bulk supply points and could not rank forecasts at the headline limit (exp 84; 2 BSPs, one scored year; provisional). See the curtailment cost theme.

What may be published

  • Never publish an NGED metered generator's time series with its name or identifier. The rule covers charts, tables, leaderboards, dashboards, reports, and issue and pull request bodies. The full rule is in CLAUDE.md (docs).
  • Use the labels from studies.anonymise: "A" to "F" for the solar farms and W1 to W3 for the wind farms. site_labels_for assigns them under a fixed seed, and it raises on a different count of generators (code).
  • Put no calendar dates on the axis of a generator's time series, and publish no output in MW. Normalise by the generator's own capacity. The dataviz skill's anonymisation section and the study skill's anonymisation section own the rest (docs).
  • Keep any plot of one farm's time series, and the cleaned series, under data/ml-experiments/. The private tables stay there (docs).
  • D1 to D22 label demand-side series, not generators in general. The glossary defines them, and substations may be named (docs).

Public generator data

  • Time series from public sources may be named. The settled output of Balancing Mechanism Units in Elexon's published data may be shown with names, in MW, on calendar dates. A study that mixes public and NGED generators anonymises the NGED ones (docs).
  • The public-data studies are the solar BMU census and the solar BMU disaggregation. (docs)

Last checked against experiment 218.