Abstract: In recent years, although the overall ambient air quality in China has improved, ozone (O₃) pollution has intensified, posing threats to both ecological environment and human health. Existing O₃ concentration prediction methods often struggle to capture complex non-linear relationships and long-term dependencies in time series data, resulting in limited prediction accuracy. To address these challenges, this study develops a Residual Long Short-Term Memory Network (ResLSTM-Ozone) model, integrated with the SHAP explainability method to analyze feature contributions, using daily data (pollutants, meteorological data, etc.) from national control stations in Chengdu from 2014 to 2023. The model performance is evaluated using the R² and root mean square error (RMSE). The results show that: (1) O₃ concentrations in Chengdu show significant seasonal variation, with a daily mean ± standard deviation of (90.9±49.2) μg/m³ from 2014 to 2023. Ozone exceedances occur frequently during spring and summer, while concentrations are lower in autumn and winter, resulting in a total of 420 exceedance days over the 10-year period. (2) O₃ concentration variations are driven by the synergistic effects of meteorological factors and air pollutants. Temperature and sunshine duration exhibit a significant positive correlation with O₃ concentration, and the exceedance rate increases markedly with rising temperatures. In contrast, relative humidity, NO₂, PM_2.5, and CO exert inhibitory effects on O₃ levels, with their contributions showing seasonal differences. (3) ResLSTM-Ozone Model optimizes gradient propagation through a residual structure, achieving excellent prediction accuracy (R²=0.81, RMSE=22.2 μg/m³). It effectively captures overall concentration trends, with prediction biases under 15% in low-to-medium concentration range (30-120 μg/m³), though some deviations exist when predicting extreme peaks (>160 μg/m³). SHAP analysis reveals that temperature and sunshine duration are dominant positive contributors (with average absolute SHAP values ranking among the top two, and their combined positive contributions accounting for 65%-70%), while pollutants such as CO, PM_2.5 have negative inhibitory effects (total contributions of approximately 30%-35%). The complementary effects of positive and negative features enhance the model′s ability to simulate complex atmospheric processes. Overall, the study indicates that O₃ pollution in Chengdu is primarily influenced by local photochemical reactions under high-temperature, strong solar radiation, and stable meteorological conditions. The constructed deep learning model accurately predicts O₃ concentrations by capturing dynamic interactions between meteorological and pollutant factors, providing a feasible method for urban O₃ pollution prediction.