Inzinion/us-ccgt-operational-benchmark

US CCGT Operational Benchmark

Measured annual performance of the US combined-cycle gas fleet - net thermal efficiency, heat rate, CO2 intensity, capacity factor and net generation from US EPA eGRID2022 (CEMS/metered) - joined to fleet structure from EIA-860, WRI and GEM, with leakage-safe ML task definitions and a deterministic cv_fold column.

Canonical record: doi.org/10.5281/zenodo.22172560 · Publisher: Inzonex

Load

from datasets import load_dataset

ds = load_dataset("Inzinion/us-ccgt-operational-benchmark", split="train")
print(ds)

Or straight into pandas:

import pandas as pd

df = pd.read_csv("https://huggingface.co/datasets/Inzinion/us-ccgt-operational-benchmark/resolve/main/us_ccgt_operational_benchmark.csv")

409 US combined-cycle gas plants · one row per EIA plant code · measured annual performance (US EPA eGRID2022, CEMS/metered) · leakage-safe ML tasks · CC BY 4.0

Each row joins a plant's measured annual performance — net thermal efficiency (median 46.2%, range 22.4–65.4%), heat rate, CO₂ output rate, capacity factor, net generation — to its fleet structure: capacity, commissioning year, HRSG and gas-turbine unit counts, duct burners, HRSG bypass capability, CHP flag, carbon capture, planned retirement (EIA-860), owner and location.

Files

File	Contents
data/us_ccgt_operational_benchmark.csv	409 plants × 30 columns (incl. cv_fold)
data/dropped_mismatched_joins.csv	21 records removed by the entity guard (see below)
data/join_verification.csv	Per-row verification of every WRI/GEM → EIA join against eGRID2022
verify_eia_joins.py	Script that produced the verification table (needs the public eGRID2022 workbook)
tasks.json	Three ML task definitions with per-task excluded-feature lists
data_dictionary.csv	Field-level documentation

Why this exists

The widely used UCI Combined Cycle Power Plant dataset contains 9,568 hourly observations of one plant. This benchmark is the complementary view: many plants, compared across the fleet on regulator-reported annual data — the question shifts 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 (21 records removed)

The upstream fleet inventory matched WRI/GEM plant records to EIA plant codes by coordinate proximity. A row is kept only if its plant_id embeds the joined EIA code (USA0000613 → 613) — the identity that guarantees the structural features and the eGRID measured values describe the same asset. This removes:

• 9 duplicate neighbour joins — two neighbouring records claiming one EIA code, duplicating the same eGRID values (e.g. a 1.4 MW hospital site next to the Brooklyn Navy Yard cogeneration plant inherited its 63.8% efficiency); the embedding record is kept.
• 12 structure/measurement entity mismatches — single records whose capacity and vintage describe a different asset than the measured values joined to them (e.g. a 4.2 MW 1993 cogen carrying the measured data of the 1,267 MW Blue Water Energy Center 5 km away; old repowered sites carrying the new plant's eGRID data). Every join was verified against eGRID2022 plant names, coordinates and nameplate capacity — the full evidence table is data/join_verification.csv (regenerable with verify_eia_joins.py).

All removals are logged with reasons in data/dropped_mismatched_joins.csv. Result: one plant = one row = one set of measured values describing one asset, so a plain row-wise split has no group leakage.

Leakage-safe tasks (tasks.json)

Several column pairs are (near-)deterministic transformations of each other. Predicting a target while leaving its twin among the features reconstructs a formula, not a model:

Task (target)	Must exclude	Why
measured_thermal_efficiency_pct	heat rate, CO₂ intensity, annual CO₂, generation	efficiency = 3412.14 / heat rate; CO₂ intensity ≈ fuel emission factor × heat rate
measured_co2_intensity_kg_per_mwh	efficiency, heat rate, annual CO₂, generation	same identities; annual CO₂ ÷ generation = intensity
measured_capacity_factor	net generation, generation, annual CO₂	CF = generation / (capacity × 8760)

Structural features (capacity, vintage, unit counts, duct burners, CHP, region, coordinates) are always safe. measured_capacity_factor is offered as an optional operational covariate for the efficiency task — utilisation genuinely drives efficiency (peaking < mid-merit < baseload), but including it makes the task partly operational rather than purely structural. Use the cv_fold column (deterministic sha256 of the EIA code) for reproducible 5-fold CV.

Caveats — read before modelling

• CHP allocation artefact. 27 plants show efficiency above 55%; 25 of them are CHP sites. eGRID allocates part of the fuel to useful heat, so apparent electrical efficiency at cogeneration plants is not comparable with power-only plants. Filter or stratify on chp.
• Annual averages, one year. All measured values are eGRID2022 annual figures; they mix load-following, seasonal and outage effects.
• turbine_class is sparse (11 of 409 plants) — it cannot carry a fleet-wide technology analysis; commissioning_year is the practical vintage proxy.
• gas_turbine_count covers 393 of 409 plants; other structural features are complete.
• Heat rates follow the US HHV reporting convention; efficiencies here are net, HHV-basis, and read lower than the LHV figures in vendor brochures.
• US-only: measured CEMS data of this kind is not openly available for most other countries.

Lineage and versions

Derived from the Global CCGT & HRSG Fleet Inventory (2,784 plants worldwide, 33 columns) maintained by Inzonex — this benchmark is its US measured-performance subset after the entity guard. A separate slim global fleet edition (2,292 rows, 10 columns, no measured fields, 2026-07-18) exists for a different purpose; it is not the parent of this file.

v1.1.0 (2026-08-30): negative numeric values (longitudes) are now written as plain numbers — v1.0.x escaped them with a leading apostrophe as an over-broad spreadsheet-formula guard, which broke numeric parsing; the entity guard extended to single rows (409 plants, was 421 — 12 structure/measurement mismatches removed with eGRID2022 evidence); croissant.json upgraded to conformant MLCommons Croissant 1.0 (RecordSet + FileObject + sha256).

Sources and attribution

US EPA eGRID2022 (measured performance) · US EIA Form 860 (unit structure) · WRI Global Power Plant Database · Global Energy Monitor GOGPT · Climate TRACE. This dataset redistributes and derives from these open sources under their open licenses; see datapackage.json for links.

Cite

See CITATION.cff. Published by Inzonex — engineering data for high-temperature plant equipment.

Citation

Canonical DOI (always resolves to the newest version):

Aheiev, D. (Inzonex) (2026). US CCGT Operational Benchmark [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22172560

@dataset{inzonex_us_ccgt_operational_benchmark,
  author    = {Aheiev, Dmytro},
  title     = {US CCGT Operational Benchmark},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22172560},
  url       = {https://doi.org/10.5281/zenodo.22172560}
}