AI Framework Estimates London Air Pollution Regulation Effects

Estimating the effects of air pollution regulations on urban public health is challenging due to non-random policy implementation and confounding factors like meteorology and socioeconomic changes. This study introduces an uncertainty-aware Bayesian deep learning framework to quantify the aggregate impact of air pollution regulations on PM2.5 concentrations in London between 2010 and 2020. The framework integrates daily PM2.5 observations, meteorological data, annual socioeconomic indicators, temporal trends, and daily regulation status for 32 policy measures. It employs a Bayesian LSTM to capture temporal dependencies, Bayesian embedding layers for temporal and regulation inputs, and a regulation status prediction branch for propensity score-based adjustment. Regulatory effects are determined by comparing observed PM2.5 with counterfactual predictions under a no-regulation scenario, with uncertainty quantified through repeated Bayesian training and bootstrap resampling. The findings indicate that London's regulations were associated with an average PM2.5 reduction of 1.88 µg/m³, representing a 12.35% relative reduction. The effects became more pronounced from 2013 onwards, peaking in 2018-2019, suggesting that sustained interventions measurably improved air quality.