What Machine Learning on 42,000 Power Plants Reveals
We built an open, machine-readable dataset of 42,000+ power stations — per-plant CO₂, fuel, climate zone and ISO 9223 atmospheric-corrosivity — and ran the numbers. This report shows what asset-level data reveals once you can slice a global fleet by emissions, environment and corrosion risk. Every figure is from our own open dataset (CC BY 4.0, DOI-published); reproduce it yourself.
Most per-plant CO₂ is modelled, not measured — and that changes everything
Source: Inzonex PowerAtlas — Per-plant CO₂ dataset (4,551 stations) (2026)
Across 4,551 power stations with an asset-level CO₂ figure, only about 15% carry independently measured emissions (US EPA GHGRP + EU ETS); the other ~85% are modelled estimates (Climate TRACE). Any AI model trained on per-plant emissions inherits that split. Storing the source on every row — measured vs modelled — is the single most important design decision for trustworthy industrial-emissions ML.
A quarter of the fleet sits in high-corrosivity environments
Source: Inzonex PowerAtlas — Per-plant ISO 9223 corrosivity (42,138 plants) (2026)
Applying the ISO 9223 informative method per plant (from climate normals, Köppen zone and coast distance) to 42,138 stations: about 27% fall in corrosivity class C4 or worse, and 8.7% in C5–CX (very high to extreme). For asset-integrity AI this is the prior nobody encodes — a model that ignores the site's environment will mis-rank where outdoor steel, cladding and insulated lines degrade fastest.
Humidity, not salt, is the dominant degradation driver
Source: Inzonex PowerAtlas — Per-plant environmental severity dataset (2026)
When you classify each plant's dominant environmental stressor, humidity / time-of-wetness leads at ~56%, thermal cycling ~19%, airborne marine salt ~13% and dust/abrasion ~12%. The takeaway for predictive-maintenance and corrosion models: the most common failure driver is the quiet one (wetness), not the dramatic one (coastal salt) — so inland 'safe' sites are routinely under-protected.
Why this matters for industrial AI
Source: Inzonex PowerAtlas — PowerAtlas — open per-plant energy & emissions data (2026)
Per-asset, machine-readable data with honest provenance is the raw material for every credible industrial-AI model — predictive maintenance, emissions accounting, corrosion-risk ranking. The lesson from building it: expose the source and confidence on every value, attach the physical environment to every asset, and treat estimates as estimates. Explore the live data at PowerAtlas, or rank your own equipment by ISO 9223 corrosivity using the open method.
Related
Sectors: Power Generation · Chemicals · Steel & Metals · Cement