[Analysis of Influencing Factors of Ozone Concentration in Xi'an Based on Explainable Machine Learning]

Based on the atmospheric pollutant and meteorological monitoring data of Xi'an from 2020 to 2024， this study uses machine learning models such as XGBoost and combines with the SHAP interpretability algorithm to quantitatively analyze the key influencing factors of ozone （O₃） concentration. The results showed that the XGBoost model had excellent prediction performance， with a determination coefficient （R2） of 0.92， mean absolute error （MAE） of 9.52 μg·m^-3， and root mean square error （RMSE） of 13.19 μg·m^-3. Meanwhile， its prediction deviation for high-concentration O₃ （>160 μg·m^-3） was significantly lower than that of other models. Analysis via the SHAP model indicated that meteorological factors played a dominant role in O₃ formation， contributing 61.7%， among which temperature， relative humidity， solar radiation， and sea level pressure were the main driving factors. Atmospheric pollutants contributed 39.3%， with nitrogen dioxide （NO₂） accounting for 73.8%. Seasonal differences in the contribution of various factors to O₃ formation showed that high temperature dominated in summer （temperature contributed 36.4%）， while NO₂ made the main contribution in winter due to increased heating emissions， with a contribution rate of 30.2%. The temperature， solar radiation， and PM_2.5 showed non-linear positive correlations with O₃， while NO₂ and relative humidity showed non-linear negative correlations with O₃. When the temperature >20℃， the NO₂ concentration <17 μg·m^-3， relative humidity <65%， solar radiation >1×10⁵ J·m^-2， and PM_2.5 concentration was in the range of 100-200 μg·m^-3， and the positive contribution to O₃ formation was significant. During the daytime， solar radiation and temperature synergistically activated photochemical reactions， NO₂ underwent photolysis to produce precursors， and these processes together drove the increase in O₃ concentration. This process corresponded to the "high-contribution interval" in SHAP scatter plots and the "peak period" in diurnal variation charts. In contrast， the positive contribution of PM_2.5 to O₃ was relatively significant around midday （11：00-15：00）， while it shifted to a negative impact during other time periods.