Characterising Multivariate Air Pollution State Evolution in an Urban Atmosphere Using Deep-Learned Baseline Representations: London

Urban air quality management has been playing a significant role for its effects on public 2 health and pollution characteristics of countries with constantly changing policies. Tradi- 3 tional approaches capture how much pollution is present but are unable to detect changes in 4 the chemical character of the atmosphere, the relationships between co-emitted species, the 5 balance of photochemical processing, and the combustion fingerprint of emission sources. 6 This study introduces a framework that identifies and diagnoses such evolutions within 7 the pollutants of the atmosphere. A chemistry-aware Variational Autoencoder is trained 8 on 19 multivariate pollution features (7 raw concentrations, 5 chemical ratios, 7 temporal 9 gradients) at London Marylebone Road (urban roadside) and North Kensington (urban 10 background) during 2015–2019, and tested on 2022–2025. A four-method ensemble frame- 11 work (VAE reconstruction error, reconstruction probability, Isolation Forest, and statistical 12 Z-score) requires ≥3 agreement to identify high-confidence departed pollution states. Per- 13 feature decomposition of the reconstruction probability diagnoses the chemical character 14 of each departure. At roadside site, 14.5% of post-COVID hours fall within departed states, 15 dominated by the CO/NOx combustion ratio (513.2) and the photostationary state proxy 16 (391.4), chemical relationships rather than individual concentrations. This indicates that 17 at the point of emission, London’s fleet modernisation and Ultra Low Emission Zone 18 (ULEZ) have changed the combustion fingerprint and photochemical equilibrium. The 19 same structural indicators are carried over during the COVID-19 lockdown; however, O3 20 rises 3.2× during the pandemic period, reflecting suppressed NO titration. Conversely, 21 at the urban background site, where the departures are driven by concentrations and 22 boundary-layer trapping (r = −0.659), the combustion fingerprint of the atmosphere is 23 invisible to detect (CO/NOx = −45.0). These findings indicate that London’s emission 24 landscape has undergone fundamental transformations over the past decade, and the 25 consequences of ULEZ and similar interventions or greater impacts of pandemic-related 26 events are non-homogeneously distributed across the relevant region.