Evaluating Emergency Emission Controls During an Unprecedented Regional Haze Event Using Causal Machine Learning

Emergency emission controls are widely implemented to mitigate severe haze, yet their effectiveness is difficult to disentangle from meteorological variability and concurrent changes in activities. Here, we develop a three-step evaluation approach integrating air quality meteorological normalization, augmented synthetic control, and machine-learning prediction to causally quantify intervention-attributable changes in local emission strength and ambient pollutant concentrations. Applied to Shijiazhuang’s unprecedented 45-day “Sharp Sword to Cut Haze” campaign in 2016, the approach revealed that local emission strengths of PM2.5, PM10, SO2, NO2, and CO were reduced by 4.3%, 4.8%, 15.8%, 5.8%, and 15.3%, respectively. Corresponding concentrations fell by 42.1 (14.8%), 47.4 (12.4%), 13.8 (16.6%), 15.8 (15.5%) μg·m–3, and 0.9 mg·m–3 (19.6%). Observation-based PM2.5 source apportionment results independently supported these estimates: contributions from vehicle emissions, steelmaking, and coal combustion were 4.1%, 11.0%, and 21.7% lower, respectively, than during the corresponding period in 2017. Despite stringent controls, PM2.5 averaged 243.2 μg·m–3 during the intervention period. Thus, stringent local controls substantially reduced emissions, but regional transport and unfavorable meteorology partly offset the resulting air quality benefits, highlighting the need for coordinated regional mitigation.