European Heat Pump Keymark Population Dataset and COP Models for Cold-Climate Building Simulation

A statistical-population treatment of the European Heat Pump KEYMARK certification corpus for air-to-water heat pumps: deduplicated test-point data, population-scale EN 14825 SCOP validation, mixed-effects population COP models with per-product uncertainty, Estonian climate-transfer data, building archetype load profiles, and standalone conversion code for dynamic building energy simulation.

Version 2 (August 2026) — release accompanying the revised method article: two anchoring modes (shape-only and backup-aware) implemented and documented in the code; capacity-interpolation Step 4; floor-aware numerical anchoring solve; corrected code constants; new validation/ scripts (out-of-sample error tables, backup-share population, anchored spread, rating-schedule check); a temporal hold-out on a 20 August 2026 Keymark snapshot (66 snapshot-unseen models; derived cohort and error data included, model not refitted) with a deposit-only replay script; IDA ICE consistency-check meter files; automated tests reproducing the worked example. It also adds an external laboratory hold-out: the frozen population COP model (no refitting) evaluated on the published test results of the WPZ Buchs heat-pump test centre (OST – Ostschweizer Fachhochschule, IES; “Prüfresultate Luft/Wasser-Wärmepumpen”, edition 29.05.2026) — 104 tested air-to-water units parsed, a 102-unit primary cohort (49 manufacturers, 1,021 measured EN 14511 full-load and EN 14825 part-load points) after the pre-committed identity screen against the Keymark training corpus. Cold-only fixed effects: MAE 0.448 (manufacturer-cluster bootstrap 95 % CI 0.391–0.521), RMSE 0.607, mean bias −0.054, R² 0.839; full-load points MAE 0.350 / bias +0.036 versus part-load points MAE 0.623 / bias −0.218; all 104 units: MAE 0.451, bias −0.064. Shipped as derived CSVs (scop_validation/wpz_*), a deposit-only replay script (validation/wpz_holdout.py), build-stage documentation and automated checks (code/test_wpz_holdout.py); only the transcribed measured values and derived errors are deposited, the WPZ PDF is not redistributed (source statement in scop_validation/SOURCES_wpz.md). Data files and fitted models are unchanged and remain frozen to the 18 March 2026 snapshot.

Contents

• models/ — Mixed-effects population COP models of the Log+InvSqrt form COP = β₂ + β₀·ln(ΔT) + β₁/√ΔT, with per-product random effects. All-climate variant (53,551 EN 14825 test points, 3,395 heat-pump groups: β₀ = −1.8528, β₁ = +18.0463, β₂ = +7.2602) and cold-only variant (11,397 Colder-zone points, 979 groups: β₀ = −2.5819, β₁ = +6.6052, β₂ = +11.7252), plus functional-form comparison tables.
• scop_validation/ — Batch EN 14825 bin-method SCOP recalculation for 26,478 valid configuration×climate×supply-temperature rows (declared vs recalculated SCOP and deviations); implementation validated against the EN 14825:2018 Annex H reference example (3.60 vs 3.61).
• training_data/ — The 53,551 deduplicated EN 14825 part-load test points (COP vs temperature lift, with supply temperature, climate zone, and group identifiers) used for model training.
• climate/ — Estonian test reference year (1990–2020) heating-season hours mapped to 1 °C bins alongside the EN 14825 Colder template, with part-load factors and population COP per bin; population SCOP shape factors under both climates.
• archetypes/ — Synthetic hourly space-heating and DHW load profiles for two Estonian single-family-house archetypes (pre-retrofit and nZEB), with parameter manifests.
• code/ — Standalone Python implementations of the EN 14825 bin-method SCOP calculator and the four-step SCOP→hourly-COP conversion (heating curve → population COP shape → α-rescaling → capacity-aware piecewise operation) for dynamic building simulation.

Provenance and rights. All quantities are derived from the publicly available Heat Pump KEYMARK certification database (https://keymark.eu/, accessed 18 March 2026 (population models and training data are frozen to that snapshot; a separate temporal hold-out uses a later snapshot retrieved 20 August 2026, from which only derived cohort and error data are included, without refitting)). The external laboratory hold-out (part of v2) additionally uses measured values transcribed from the publicly published test-result list of the WPZ Buchs heat-pump test centre (OST; edition 29 May 2026, see scop_validation/SOURCES_wpz.md); only the transcribed factual measurements and the derived errors are deposited, not the source document. The deposit is not a verbatim copy of that registry: it contains the authors’ deduplicated, harmonised, quality-screened, and computed compilation in the authors’ own structure. Declared certificate values embedded in the derived tables are factual technical data published by the certification scheme for market transparency; any database rights of the scheme owner in the original registry are unaffected by this deposit.

Licensing. Derived data (CSV/JSON): CC BY 4.0, applying to the authors’ compilation, structure, and computed content. Code (code/): MIT License.

Funding. Estonian Research Council grant PRG2154; Estonian Centre of Excellence in Energy Efficiency, ENER (grant TK230), funded by the Estonian Ministry of Education and Research; research infrastructure “Energy efficiency and renewable energy research infrastructure” (TARISTU24-TK12).

A peer-reviewed journal article by the authors describes the methodology and findings in detail; its reference will be linked in this record once available. Please cite this deposit via its concept DOI: 10.5281/zenodo.20627396 (resolves to the latest version).