China Provincial Grid vs National Average: The Factor Your Auditor Will Question

Your supplier in Yunnan runs a smelter on hydropower. Your supplier in Inner Mongolia runs the same process on a coal-heavy grid. Your study lists one input for both: “China electricity.” The verifier opens the file, finds a national-average grid factor behind that input, and asks the question you skipped — which province? That single substitution can move the cradle-to-gate footprint more than any other line in the bill of materials, and it is the line most spreadsheets get wrong by default.

China is not one grid. HiQLCD v1.4.0 covers all 31 provinces and municipalities at provincial precision — 4,862 China energy datasets — because a coal-dominant northern grid and a hydro-heavy south-western grid deliver electricity at sharply different carbon intensities per kWh. A national average is the weighted blend of the two. For a product made in one known province, that blend isn’t a representativeness call you can defend. It’s the wrong number for the goal and scope, dressed up as a defensible one.

What a national average hides

A grid factor is the GWP100 result for one kWh of supplied electricity, and it is geography all the way down. Northern coal grids and south-western hydro grids sit at opposite ends of the same country’s spectrum. Average them and you get a figure that describes no real site — least of all the one your supplier operates.

The damage is structural, not cosmetic. Electricity is rarely a footnote in a metals or chemicals product system; more often it is the dominant contributor to cradle-to-gate GWP. So the province you assign to “China electricity” can swing the result more than the choice of any raw material. A verifier who knows the field knows this, which is why the first question about a China-sited product is almost never about the steel. It’s: which grid, and what year.

Many international databases can’t answer that. They offer a China national average, or a generic electricity dataset transplanted from another region and relabeled. Either way, the geographic provenance the verifier needs is gone before the calculation starts, and you inherit a number you can’t defend on its own terms.

None of this makes a national average useless. Where the goal and scope are genuinely national — mixed sourcing, unspecified site, country-level reporting — the blend is the right and defensible figure. The error is reaching for it when you already know the single site’s province.

Geographic is a DQI dimension, not a rounding error

The Data Quality Indicator — built on the Pedigree-Matrix lineage practitioners already know — scores every candidate dataset across five dimensions, in this order: Temporal, Geographic, Technology, Completeness, Reliability. Geographic measures how well a dataset’s region matches the process under study. A national average applied to a single-province product scores poorly on exactly that dimension — and it is precisely the substitution that throws the dimension away while leaving the rest of the file looking complete.

That’s the trap. The number looks finished. The Temporal field has a year, the GWP column has a value, the source has a URL. Nothing on the surface flags a weak Geographic match. The mismatch surfaces only when a verifier walks the trail back — or when a tool scores Geographic explicitly and shows you the gap before you file.

For a single-province product whose site you know, a national average is not a smaller error than a wrong material. It is often the largest one in the file.

What Cortex does with “China electricity”

Type “China electricity” into Cortex and it won’t quietly reach for a national average. It asks one thing first: which province, which year. That is the whole intervention, and it is deliberate. The ambiguity in your input is the ambiguity a verifier will probe; Cortex surfaces it while you can still answer cheaply, not three weeks later in review.

Answer “Yunnan, 2023” and Cortex returns the region-specific factor from HiQLCD’s provincial coverage — one of the 4,862 China energy datasets — with the DQI scored across all five dimensions. Geographic now reflects a real provincial match, not a national blend. The candidate carries its GWP100 value, unit, region, system model, and source URL. The reasoning chain records the question Cortex asked, the answer you gave, and why this dataset won. When your verifier asks “which grid,” the answer is already in the file — written down at the moment of the decision, not reconstructed after the fact.

Cortex pauses elsewhere for the same reason. When the closest match needs a proxy substitution, when the same material returns factors that differ by more than 2× across databases, when a dataset is restricted and the GWP can’t be filled silently, when the system-model match is ambiguous — the run stops and the choice returns to you. The practitioner decides; the decision is recorded in the reasoning chain. Automation handles the parallel search across fourteen LCA databases — HiQLCD, Ecoinvent, EF, CarbonMinds, and others — but it doesn’t make the calls an auditor would later question.

Province first, then defend

The discipline is easy to state and easy to skip under deadline. Before you accept any “China electricity” factor, fix the province and the year. The provincial coverage exists — all 31 provinces and municipalities, at provincial precision — so when the site is known, the only reason to ship a national average is that nobody asked the question. Cortex asks it, scores Geographic, and writes the answer into a reasoning chain you can hand to a verifier.

For a single-province product, a national average is a defensible-looking number that doesn’t survive its own audit. A provincial factor with Geographic scored is a number you can stand behind, line by line. The difference costs one clarifying question at the start. It costs a great deal more at the end.

Ask Cortex about a China-sited input at cortex.hiq.earth/chat.