Published August 30, 2026 | Version 1.1.0

US CCGT Operational Benchmark: measured efficiency, CO2 intensity and utilisation of the US combined-cycle fleet

Authors/Creators

  • 1. Inzonex

Description

409 US combined-cycle gas (CCGT) plants, one row per EIA plant code, with measured annual performance from US EPA eGRID2022 (CEMS/metered): net thermal efficiency (median 46.2%, range 22.4-65.4%), heat rate, CO2 output rate, capacity factor and net generation - joined to fleet structure: capacity, commissioning year, HRSG and gas-turbine unit counts, duct burners, HRSG bypass capability, CHP flag, carbon capture, planned retirement (US EIA Form 860), owner and location.

Why it exists. The widely used UCI Combined Cycle Power Plant dataset contains 9,568 hourly observations of a single plant. This benchmark is the complementary view - many plants compared across the fleet on regulator-reported annual data - shifting the question from "how does ambient temperature move one plant's output" to "why is one plant more efficient than another" (vintage, technology class, utilisation, cogeneration).

Entity guard. The upstream fleet inventory matched plant records to EIA plant codes by coordinate proximity; a row is kept only if its identifier embeds the joined EIA code, guaranteeing that structural features and measured values describe the same asset. This removes 9 duplicate neighbour joins and 12 single records whose capacity/vintage describe a different asset than the joined eGRID data (e.g. a 4.2 MW 1993 cogen carrying the measured values of a 1,267 MW plant 5 km away). Every join was verified against eGRID2022 names, coordinates and nameplate capacity - evidence in join_verification.csv, all 21 removals logged in dropped_mismatched_joins.csv. One plant = one row = one set of measured values.

Leakage-safe ML tasks. tasks.json defines three regression targets (thermal efficiency, CO2 intensity, capacity factor) each with an explicit excluded-feature list, because several columns are deterministic transformations of one another (efficiency = 3412.14 / heat rate; CO2 intensity is approximately fuel emission factor x heat rate; capacity factor = generation / (capacity x 8760)). A deterministic cv_fold column supports reproducible 5-fold cross-validation.

Caveats. 25 of the 27 plants above 55% apparent efficiency are CHP sites, where eGRID allocates part of the fuel to useful heat - filter or stratify on the chp flag. Values are annual averages for a single year (eGRID2022), efficiencies are net and HHV-basis (lower than LHV vendor figures), turbine_class is populated for only 11 of 409 plants, and coverage is US-only because measured CEMS data of this kind is not openly available for most other countries.

v1.1.0: negative numeric values (longitudes) are now plain numbers (v1.0.x escaped them with a leading apostrophe as an over-broad spreadsheet-formula guard, breaking numeric parsing); the entity guard extended to single rows (409 plants, was 421); croissant.json upgraded to conformant MLCommons Croissant 1.0.

Sources: US EPA eGRID2022; US EIA Form 860; WRI Global Power Plant Database; Global Energy Monitor GOGPT; Climate TRACE. CC BY 4.0.

Files

croissant.json

Files (191.6 kB)

Name Size Download all
md5:b1998d30f031c8700365ad89687f00d4
8.7 kB Download
md5:36196330b3abf80d848660c65fd891dc
391 Bytes Download
md5:0a0d065d45821fc08a93f9978d1853e8
26.5 kB Preview Download
md5:90734a9c98568b4afd7f2d37863e2d62
2.2 kB Preview Download
md5:bb48161b3d0ef8c1caffd63fd0f31425
8.0 kB Preview Download
md5:7d9a350c30bbcfc737ceffd8e3702828
674 Bytes Preview Download
md5:3719632d91ef2f8f03ffb1c777951767
3.3 kB Preview Download
md5:cc5162fb28fc342526b40e37bd6a904b
41.7 kB Preview Download
md5:f02a7a983e2686b79576a94c3bc0dea2
657 Bytes Download
md5:9883460cc6fddf724fa01b12582f1abf
6.4 kB Preview Download
md5:42f542f5f76ca06b727eb181916cc5eb
2.8 kB Preview Download
md5:0325c1249e59b034e8437a29ea8dd6f4
86.1 kB Preview Download
md5:fa45fee423354e73417770c2a7aca757
4.2 kB Download

Additional details