A hybrid deep learning model for O₃ forecasting and explaining in the Yangtze River Delta Region of China

The advancement of explainable artificial intelligence (XAI) has emerged as a pivotal tool for unraveling the intrinsic mechanisms of deep learning-based models for predicting air quality. The Layer-wise Relevance Propagation (LRP) method enables the quantification of temporally resolved feature contributions in deep learning (DL) models predicting ozone (O₃) concentrations. In this study, a Random Forest-corrected (RF-corrected) model was employed to generate 9-km gridded O₃ data for the Yangtze River Delta (YRD) region from 2020 to 2023, filling gaps in unmonitored areas. These gridded data, combined with WRF-modeled meteorological parameters and ground observation data, served as inputs for an attention-based sequence-to-sequence (seq2seq) model to predict O₃ concentrations at 24-h (24 h), 48-h (48 h), and 72-h (72 h). For the 24 h predictions, the model demonstrates robust performance, with R of 0.85 (test set) and 0.89 (validation set), alongside RMSE of 19.81 μg/m³ and 19.22 μg/m³, respectively. LRP analysis revealed annual average contributions of 24.6 % from gridded features, 38.2 % from meteorological features, and 17.0 % from pollutant features to O₃ predictions. During 13:00-15:00, contributions increased to 30.0 %, 35.0 %, and 15.3 %, respectively, aligning with midday photochemical O₃ production and regional transport under favorable meteorological conditions. Further analysis integrating Potential Source Contribution Function (PSCF) values and industrial zone locations near monitoring sites demonstrated that gridded feature contributions effectively explain the influence of O₃ sources from different wind directions. These findings highlight that deep learning-derived feature insights, combined with spatial and chemical context, provide novel perspectives for identifying emission sources and transport mechanisms driving O₃ variability.