Hierarchical deep learning framework for ozone forecasting and source apportionment: A China-wide analysis

Abstract

China's deteriorating ozone (O₃) pollution has garnered increasing attention, necessitating accurate forecasting and source attribution for effective mitigation. In this study, we developed a hierarchical graph neural model with adaptive multi-scale edge weights, integrating both city- and province-level spatial interactions to dynamically capture cross-regional transport and chemical processes. The proposed model achieved superior forecast performance, with R² = 0.98, RMSE = 3.41 μg/m³, MAE = 2.71 μg/m³, based on the national-average O₃ concentrations across 326 cities. Channel attention analysis demonstrated marked regional heterogeneity in the dominant factors affecting O₃ forecasts, with systematic variations observed during the polluted season. Moreover, we introduced an edge-weight-based framework for O₃ source apportionment, which quantifies cross-regional influences and captures how meteorology and chemical species jointly shape pollution networks. This study presents a novel framework for advancing O₃ prediction and for developing effective, regionally coordinated control strategies.