A Double Machine Learning Approach for Policy Evaluation Using Meteorologically Fixed Air Pollutant Time Series as an Emission Proxy

Quantifying the effectiveness of air quality policies requires controlling for meteorological confounding. Traditionally, this has relied on scenario-based chemical transport model simulations, yet their accuracy is constrained by uncertainties in emission inventories, particularly their inability to represent abrupt emission changes induced by short-term interventions. Statistical “de-weathering” methods using real-world observations offer an alternative, and machine learning (ML)-based meteorological normalization has become increasingly popular. However, existing approaches normalize pollutant time series to an assumed “average” meteorology, often generating inconsistent normalized conditions and reducing sensitivity to short-term emission changes. We introduce a double-ML approach to evaluate policy-driven air quality responses. The first ML model, MetFix-normalization, predicts pollutant levels under consistent meteorological states. Using London roadside NO2 observations, MetFix-normalization shows stronger agreement with traffic activity indicators and with prescribed emission-perturbation scenarios than the widely used normalization technique. MetFix-normalized NO2 is then used as an emission proxy in a second ML model to predict business-as-usual NO2 levels under actual meteorology with counterfactual emissions, enabling evaluation of the impacts of the COVID-19 lockdown and mitigation policy. An automated version (auto-MetFix) was developed to support near-real-time tracking of emission. This double-ML provides a data-driven approach for assessing emission–air quality relationships under changing emissions.