| English | 中文 |
Station observations support concentration comparisons, model evaluation and exposure research. Distinguish original hourly publication records, city summaries and processed daily products. Operations are in the skill, coverage in the source table, and adjusted daily values in the homogenized method.
The example follows national stations → CNEMC publication → Quotsoft archive → selected-date acquisition → long table. The provider description identifies the upstream source. This is a third-party historical record, not an official historical API, instrument signal or fully audited operating database. Hourly PM2.5 is mass concentration in µg/m³; PM2.5_24h and AQI are different quantities. Complete per-record calibration/maintenance/review evidence is absent.
| Error source | Research handling |
|---|---|
| Flow, inlet condition, conditioning, calibration and instrument changes | Retain available source flags and maintenance evidence; hourly CSV cannot reconstruct minute-level validity or instrument correction |
| Relocation, renaming, surroundings and point representativeness | Use dated station history and explicit matching; a point does not represent every person or the full grid cell |
| Collection failure, delay, revisions and duplicate retrieval | Separate absent files, hours, station values and quality exclusions; failure is not zero concentration |
| Day/night or weather-dependent availability | Report missing times/seasons and sample composition before/after completeness screening |
| Duplicate hours, moving averages, imputation and outlier removal | Count unique observed hours and preserve original/processed flags; high concentration alone is not an error |
Measurement considerations follow HJ 817—2018. Bai et al. (2020), sections 2–3 discuss discontinuities and processing in that study; these are possible mechanisms, not measured station error sizes.
The 2018 GB 3095 amendment, effective 2019-01-01, expresses particles at actual temperature/pressure and gases at reference conditions. Check retrospective conversion before trend analysis; do not convert by year alone or interpret a definition change as emission reduction. Temperature/pressure evidence and an explicit formula are needed. See the amendment and ministry clarification.
This source table preserves the review dated 2026-10-26.
| Basis | Requirement | Scope |
|---|---|---|
| GB 3095—2012 section 6.4/table 4 | At least 20 hourly values or sampling hours for PM2.5 daily mean | Historical evaluation/research adopting that edition with applicable amendments/evaluation rules |
| GB 3095—2026 chapter 6/table 4, HJ 663—2026 section 6.1.1 | Minimum remains 20 hours; evaluation statistics follow GB 3095 | Effective 2026-03-01, not the edition effective in 2024 |
| Bai et al. (2020) section 2 | More than 12 valid hours, hence at least 13 | That paper’s US embassy/consulate independent validation sample only, not all data or the national criterion |
The announcement states effective dates. Regulatory evaluation follows applicable rules; reproduction follows the exact paper/data object; longitudinal research predetermines a consistent policy and explains differences from regulation.
n_valid_hours/24; daylight-saving days elsewhere need actual day length.n_valid_hours >= 20, averaging valid hours rather than dividing by 24. Otherwise retain daily_mean=NA, daily_valid=false and insufficient_valid_hours.Synthetic boundaries: 19 valid unique hours fail; 20 pass; 20 rows with one duplicate give at most 19 hours; 24 missing rows give zero. Counts do not repair missing source/time/instrument evidence.
Retain n_valid_hours, n_expected_hours, coverage_fraction, daily_valid, exclusion_reason, quality_policy_id, source_qc_status and imputed_hours. Monthly/annual summaries need their own valid-day/month coverage; pooling available hours does not establish a compliant annual mean.
Flag negatives, out-of-range, constant and abrupt records separately; review evidence before removing pollution extremes. Imputation, relocation merging and homogenization produce separate products. Reconstructed values cannot independently validate their own reconstruction.
Before ERA5/CAMS comparison, align UTC/local days, averaging, concentration state, surface definition and spatial support. Model values do not become station observations. Choose station/time validation splits and fit normalization, imputation and bias correction on training data only. See preparation.