Skip to content

Substation information fields

Four sources describe a substation: the metadata table, NGED's hierarchy files, the effective-capacity table, and the estimates of solar capacity that experiments have built. This page lists what each source holds, where it sits, and which fields have misled an experiment. The power readings themselves are on the NGED power telemetry page, and the paths are on the disk page.

The metadata table

  • data/NGED/metadata.parquet holds one row per time series, under the TimeSeriesMetadata contract. The field descriptions in packages/contracts/src/contracts/power_schemas.py own the meaning of each column (code).
  • The trial area has 33 rows: 16 Primaries, 2 BSPs, 2 GSPs, and 13 customer connections. Of the 13 connections, 7 are typed PV, 3 Wind, and 1 each BESS, Biofuel, and Other (Generation) (disk).
  • Only the 16 Primaries carry an area polygon (area_wkt), and every customer row has a null polygon (disk).
  • information is a free-text note, filled for 5 of the 33 rows. The notes are a snapshot from July 2026, because the earlier raw files carry no value, so a note cannot say when a fault began (exp 182; 33 rows; provisional).

The hierarchy files

  • Four NGED files and a README sit in data/NGED/substation_hierarchy_2026-10-15/. They are Substations.csv, Primary Substation Interconnections.csv, Monitor Direction.csv, and Peak Loads.xlsx. They are private (disk).
  • Substations.csv gives each substation's BSP and GSP. The trial area has two GSPs, each with one BSP (disk).
  • Primary Substation Interconnections.csv lists pairs of Primaries that can be connected. Every trial-area Primary has at least one pair, so the topology looks complete (data sources; docs).
  • Monitor Direction.csv gives each series' meter type and flow direction. One solar farm's substation number differs from its metadata row, and the other sources agree with the metadata (docs).
  • Peak Loads.xlsx disagrees with the telemetry at several Primaries, so experiments take the 99th percentile of observed power as the capacity of a series instead (docs).
  • The Primaries beneath a BSP do not sum to the BSP's reading. Primaries that report only apparent power, and the generators and unmetered load beneath the BSP, stop the sum from closing. Adding every child Primary to a bottom-up sum did not improve the BSP forecast (exp 161, 204; 2 BSPs, one scored year; provisional).

The effective-capacity table

  • data/effective_capacity/ holds one value per series: the 99th percentile of the absolute power over the series' whole observed history, in MW. It is not a nameplate, firm, or connection rating (code).
  • Scores in points of capacity divide by this value. See the glossary (code).
  • The value is built from the series' own output, so a farm's power divided by it near 1.0 is not an independent level (exp 179, 183; 6 farms; provisional).

Estimates of solar capacity behind a Primary

  • No source tells us the solar capacity behind a Primary, so experiments estimate it from the Primary's power. The estimates are AC capacity in MW, taken as the 99th percentile of the solar output over daylight half-hours (exp 129; 5 Primaries; provisional).
  • Experiment 45 built weekly capacity "vintages", one block per origin Monday, and the frozen variant uses data before the origin only. A cut earlier than a Primary's first vintage receives a later vintage's level, which is hindsight. The rule is in the studies.dp_pinned_capacity docstring (code).
  • The latest week of a real-time vintage has 3.6 times the path error of the same week in the final fit. The pinned-capacity code therefore holds the median of the last 13 weeks (exp 45; 8 Primaries with injected paths; provisional).
  • Experiment 129 gives one capacity per season-year at D5, D10, D11, D14, and D16, with a range. For 2025/26 the estimate is 3.3 to 4.5 MW per Primary at the five tested, and the range is a spread over test scenarios, not a calibrated interval (exp 129; 5 Primaries; provisional).
  • Reading the capacity from the Primary's own power did not improve the forecast. An as-of capacity scored 0.110 points of capacity worse than the pinned vintage on the sunniest third of daylight half-hours (exp 139; 5 Primaries, one scored year; provisional).
  • A curve of irradiance with no capacity estimate matched the pinned capacity at five of six Primaries and beat it at D5. B minus A on sunny daylight half-hours was +0.025 [-0.252, +0.346] over the five and -0.969 [-1.708, -0.176] at D5, so the D5 pin is a suspect (exp 350; 6 Primaries, one scored year; provisional).
  • A monthly capacity path recovered less of an injected farm than a yearly one, 0.67 to 0.78 against 0.75 to 0.87 (exp 157; 6 Primaries, 5 injected farms; provisional).
  • Refreshing the capacity table once per season-year lowered the error by 0.018 points of capacity, and a weekly refresh made no distinguishable difference (exp 78; the fold A row set; small gain; provisional).
  • The one dated list of installations we hold, the Feed-in Tariff register, ends on 2021-03-24. No capacity step after that date can be matched to an installation (exp 215; 6 Primaries; provisional).
  • The published estimates are on the embedded solar capacity study (docs).

Last checked against experiment 218.