Physicochemically Motivated Decomposition for Long-Sequence Urban O3 Forecasting: A Cross-Method Comparison

Surface ozone (O3) forms via nonlinear chemistry and is highly non-stationary, complicating 168-h forecasts. Many models omit key physical drivers, mix time scales, and degrade quickly with lead time. We propose a decomposition-ensemble framework that fuses surface observations with NWP predictors (e.g., solar radiation, PBL height). To prevent data leakage, decomposition (STL, GAM, CEEMDAN, VMD) was performed exclusively on the training and validation data to isolate the annual, diurnal, and residual components. We then train LSTM and Transformer models on each component using both direct and incremental (ΔY) learning. Across Beijing, Shanghai, Guangzhou, and Chengdu, decomposition lowers the error bound and curbs horizon-wise error growth. Among the tested strategies, VMD (K = 2) and period-based merging (for CEEMDAN) robustly isolate the 24-h diurnal mode, limit cross-scale leakage, and provide the best balance of separability, interpretability, and forecast skill. The results further reveal pronounced heterogeneity in optimal model architectures across temporal scales and city-specific regimes, indicating that a single universal model is insufficient to jointly optimize short-term accuracy and long-term stability. Overall, this study provides a scenario-adaptive forecasting paradigm that offers practical guidance for O3 pollution early warning and decision support.