ADR: comparison basis of regional electricity backgrounds¶

Status: accepted. Code: domains/assessment/lca/comparability.py, FactorScope in schemas.py, backgrounds.py (packaged scopes), the method-mix check in report.py. Builds on lci-scenarios.md and the grid electricity profiles.

Context¶

The packaged grid profiles give sourced factors for many geographies, but they measure different things. The process library's ASU numbers (UK 66.7, NL 88.3, DE 104.7 kg CO2e per t GOX, exergy allocation) put DESNZ (electricity at the consumer, with T&D and well-to-tank) beside EEA ENER038 (gross generation, direct emissions only). They are screening sensitivities, not like-for-like country footprints, and nothing here has independent industrial validation. Core used to tell two factors apart only by parsing the free-text method of their source, which cannot express most of the differences below.

The comparison dimensions¶

Two backgrounds may be set side by side as a like-for-like pair only if they agree on every one of these. Each is a field of the optional scope of a source or a factor.

dimension	values	why it matters	severity if it differs
`basis`	`direct` (emissions at the plant), `direct+upstream` (plus fuel supply chain, well-to-tank), `life-cycle` (plus infrastructure, manufacturing)	upstream and infrastructure can be tens of percent of a low-carbon grid	hard
`gases`	`co2`, `co2e`	CO2 only omits CH4 and N2O	hard
`accounting`	`generation` (territorial: the geography's own generation), `consumption` (trade-adjusted)	imports and exports move the figure of a country in either direction	hard
`delivery`	`generation` (per kWh generated), `consumer` (at the point of use, T&D losses included)	a loss uplift of a few percent, to be stated	soft
`gwp`	`AR4`, `AR5`, `AR6`, `AR4/AR5`, `mixed`	only for CO2e; changes the CH4 and N2O weights	soft
`mix`	`grid-average`, `residual-mix`	a residual mix excludes what contracts claimed	soft
data year	the factor's `year` (see below)	a grid decarbonises by several percent a year	soft

Not a dimension yet: gross versus net generation. eGRID divides by net generation, EEA and Ember by gross generation (own use of the power plants included); all are delivery: generation. The difference is a few percent; today other dimensions (gwp, basis) already keep such pairs apart, and a future pair that differs only in this way must state it in note.

Hard differences measure a different quantity: the totals are not comparable. Soft differences are biases that a stated uplift or caveat can bound: the pair is a proxy.

Data year versus publication year. year keeps its meaning: the reference year a factor file gives, used as given. Several sources date a file by its publication (desnz-ghg-2026:electricity-gb is year 2026, the year of the 2026 set). A factor may also declare, optionally and additively, data_year (the year the data describe) and publication_year (the year the source published or versioned them); data_year may not lie after publication_year (decision 6).

Migration: nothing changes for a factor without data_year; its records keep their bytes (publication_year is not written into records). Declaring a data_year on a factor a scenario links changes that scenario's record (and possibly its verdict): rerun cheme assess lca-scenarios --write then. desnz-ghg-2026 declares only publication_year, so projects on it need no rerun for this.

Decisions¶

Structured scope, optional and additive. Source.scope and Factor.scope (a FactorScope, all fields optional). A factor's scope overrides its source's field by field. An unset field is not declared; core never assumes it equal, with one exception: two factors citing the same source share its methodology, so a dimension both leave undeclared is not a difference between them (EEA NL against EEA DE). A project's copy of a packaged source next to its include keeps the packaged scope (the copy's own fields win). Existing files load and compute as before.
Packaged profiles keep their bytes. A packaged profile edit changes its fingerprint in every project's lca/manifest.json and scenario records, so the scopes of the five existing profiles live in data/background_scopes/<profile id>.yaml and are applied when the profile loads. The profile files and the manifest are unchanged (the manifest does not record scope). A scope written inside a profile file wins over the sidecar, field by field. Scopes were read from each profile's own notes; where a profile does not state a dimension it stays undeclared (EEA's GWP basis; DESNZ's accounting). The sidecar is not fingerprinted in the manifest (that is what keeps it stable): a change to it ships with a new software version and a changelog entry, and an entry that matches nothing in its profile is an error. The verdict looks at replaced backgrounds, the electricity procurement (contractual, residual-mix, on-site make a proxy) and gaps; a different design case changes the foreground, which the record states separately (kind: design-case).
Method-mix reads the scope. cheme.lca-method-mix warns, per category, when the contributions differ in basis, gases or (for CO2e) gwp, as declared. Contributions without a declared gwp fall back to the former reading of the source's method text, so projects without scopes behave as before (EEA, whose scope leaves gwp undeclared, is therefore read as AR5 from its method text here, while the verdict treats it as undeclared). A CO2-only factor has no GWP basis: it counts as the method CO2 only, so it is still flagged beside a CO2e factor whose GWP basis is known from either source, and gases tells it apart where both declare it.
Scenario comparability verdict. For every scenario, against the baseline:
like-for-like: every dimension of every replaced background is declared and equal, no exchange is dropped, no background is declared a proxy, no foreign geography is kept;
proxy: no hard difference, but a soft one, an undeclared dimension, a proxy: declaration, or a kept background of another geography;
not comparable: a hard difference, or an exchange the baseline characterises is a gap (the total does not stand).

Declaring proxy: discloses a difference, it never makes a hard one comparable. A scenario that replaces no background keeps the baseline's and is like-for-like on the background. The baseline's own entry is baseline. 5. Where it surfaces. lca/scenarios/<id>.json gets comparability (verdict, reasons, compared: each replaced pair with its differing dimensions), and each linked exchange its scope (null: not declared). results.md gets a comparability row, one line per scenario and marks a not comparable scenario's totals "(not comparable)". cheme assess lca --json and GET /api/lca get comparability (per scenario) and the sources' scope, additively; their older not_comparable (and a record's impacts.comparable) keep their meaning: gaps only. 6. Data year and publication year. Factor.data_year and Factor.publication_year are optional; year is never reinterpreted. The year dimension of the verdict compares data_year where a factor declares it and year otherwise, and says which (compared[].years, and [data_year / year] in the reason) whenever a data year takes part; publication_year is metadata and never compared. Packaged profiles declare years in the sidecar (years:, factor id glob -> data_year / publication_year; every matching entry counts, two giving one key different values are an error, the profile file wins). desnz-ghg-2026 declares publication_year: 2026 (its year is the publication year); its data year (2025 by the DESNZ methodology, section 3.8, see lca.md) is not declared until checked against the paper. The checks of time representativeness (cheme.lca-reference-year, cheme.lca-scenario-year, the dataset's temporal data quality) still read year, on purpose until a follow-up moves them to data_year where declared.

Consequences¶

Records and results.md of projects with scenarios gain these fields: one --write of cheme assess lca --write and cheme assess lca-scenarios --write after upgrading. Manifests of projects without scenarios do not change.
A project's own factors without scope give proxy, never like-for-like: declaring a scope is the way to a like-for-like set.
The ASU's NL and DE scenarios (DESNZ UK baseline, EEA 2024) differ in basis and are therefore not comparable with the baseline; a like-for-like set needs one declared basis (for example the Ember life-cycle series for all countries, or generation-based factors for all with a labelled loss uplift). That decision belongs to the process library.
A like-for-like set of grid factors for several countries is built from the user's own copy of Ember's long-format data with scripts/build_grid_profile.py: one basis for all (basis: direct, gases: co2, accounting: generation, delivery: generation, mix: grid-average), data_year set, the release, licence (CC BY 4.0, as Ember's page states), attribution and the SHA-256 of the copy recorded. The package does not ship the result: the script writes a project file, so no third-party data is redistributed beyond what the project chooses, and the packaged profiles keep their bytes. The set is CO2 only and direct: it is not comparable with the packaged life-cycle ember-2025, with DESNZ and with the CO2e EEA series (the verdict says which dimension differs). Chosen as the conservative basis because it is the one Ember's current long-format file states; a CO2e or life-cycle series needs a file that carries it.
Not done: other sources for the generator (EEA ENER038, eGRID, UBA, CBS: files and terms not checked for it), consumption-based variants (T&D losses, trade balance), the UBA PDF series (CO2e, upstream; awaiting written clarification from UBA); the time-representativeness checks still read year (decision 6).