Interpretable Data-Driven Ozone Prediction Using Statistical Diagnostics, XGBoost, SHAP and Temporal Fusion Transformers

This study develops an interpretable, data-driven framework for forecasting daily MDA8 ozone levels in the Beijing–Tianjin–Hebei (BTH) region, integrating statistical diagnostics, XGBoost-based SHAP feature interpretation, and the Temporal Fusion Transformer (TFT). Using two years of pollutant and meteorological data from 56 monitoring stations, we identify a dual temporal structure: ozone, temperature, and pressure follow non-stationary annual cycles, while eight other variables show stationary, autocorrelated short-term fluctuations. SHAP analysis reveals that temperature, followed by relative humidity, NO2, particulate matter, and pressure, are key predictors, in line with photochemical mechanisms. A hierarchical ablation experiment shows that multivariate models outperform bivariate ones, and meteorological variables improve predictions more than primary pollutants. The inclusion of five pollutant variables worsens performance due to multicollinearity. The XGBoost-TFT hybrid model, which compresses covariates into a single index, achieves the best performance (median R2 = 0.686), outperforming raw-input models. These results validate the framework’s interpretability and alignment with photochemical mechanisms.