Deep Learning-Based Medium- and Long-Term Ozone Forecasting for Megacity Shanghai, China

Accurate ozone concentration forecasting is crucial for effective air quality management and environmental protection. However, most of the existing forecasting models focus on short-term (24–72 h) forecasts and lack validation for extended horizons (weeks to months). Current understanding of how deep learning models bridge the temporal scale gap between short-term ozone fluctuations and long-term concentration trends is limited. This study presents a comprehensive analysis of the ozone-meteorological relationships and multi-horizon forecasting in Shanghai, a megacity in China during 2014–2021. We systematically evaluate six state-of-the-art deep learning models (Transformer, Informer, FEDformer, Autoformer, NS-Transformer, and iTransformer) for ozone concentrations forecasting across six temporal scales (96–1440 h). The correlation analysis identifies a significant positive correlation between ozone concentration and temperature and a significant negative correlation with relative humidity. Comparative model evaluation reveals distinct temporal performance patterns: Informer achieves superior accuracy for sub-two-week forecasts, while NS-Transformer demonstrates optimal performance for extended forecasting windows beyond 2 weeks. The incorporation of meteorological covariates as dynamic exogenous variables enhances the accuracy of the model, especially the monthly recalibration provides additional robustness. These results demonstrate the necessity of selecting temporal-scale-adapted models for ozone forecasting and the crucial role of meteorological characteristic engineering in long-term air quality forecasting. The physics-constrained machine learning model proposed in this study provides a scientific basis for sub-seasonal ozone prediction on a 14–60 day horizon.