Predicting Short-Term Air Quality Index in the Beijing–Tianjin–Hebei Urban Agglomeration: A Comparative Assessment of Linear, Ensemble, and Recurrent Forecasting Models

The Beijing–Tianjin–Hebei (BTH) region faces complex air pollution driven by alternating particulate matter (PM) and ozone (O3) dominance, regional transport, topography, and meteorology. This study develops a hybrid framework integrating air quality index (AQI) records, pollutants, meteorological variables, and MEIC emissions from the BTH region (2018–2025) to capture spatiotemporal evolution and short-term predictability. Results show a seasonal AQI cycle (winter/spring highs, summer/autumn lows) with a summer PM–O3 seesaw. Spatially, three zones were identified: the northern and coastal ecological barrier zone, the central compound-pollution plain zone, and the southern heavy-industrial zone. Random Forest identifies PM as the dominant AQI compositional contributor, with visibility, dew point, humidity, and MEIC emissions (particulates, NH3, organics) as key correlates. Forecast evaluation reveals progressive improvement: ARMA captures linear baselines (R2 = 0.318, MAPE = 33.26%), XGBoost improves statistical prediction by incorporating nonlinear feature interactions and lagged meteorology (R2 = 0.567, MAPE = 24.81%), and LSTM shows the strongest statistical predictive performance (R2 = 0.613, MAPE = 22.32%). The improvement of LSTM over XGBoost is incremental and reflects enhanced data-driven representation of short-term AQI–meteorology temporal dependence, rather than identification of physical pollution mechanisms. Regional disparities persist, with higher predictability in the southern heavy-industrial zone and lower accuracy in the northern and coastal ecological barrier zone affected by intermittent dust intrusions and frontal passages. Overall, the results suggest that LSTM may support data-driven short-term AQI warning, but source-oriented mitigation still requires process-based tools, such as chemical-transport or source-apportionment models.