Abstract
Abstract Previous machine learning approaches for surface ozone (O 3 ) forecasting are often limited to site‐specific applications and struggle to capture regional pollution dynamics. Here, we introduce an inverted Transformer (iTransformer), a time‐series model that integrates hourly satellite‐derived surface O 3 with key meteorological and environmental drivers. The model is applied to eastern China and demonstrates strong skill in 72‐hr surface O 3 forecasting, achieving an overall correlation of 0.86 (mean bias = 0.29 μg/m 3 ). It effectively captures spatial patterns, diurnal variability, and high‐ozone episodes, maintaining stable performance over time. During pollution events, the model reproduces both localized and large‐scale ozone enhancements with high fidelity. Forecast biases remain low, particularly at shorter lead times, indicating robust predictive skill. Overall, the results highlight the ability of advanced deep learning architectures to bridge satellite observations and air quality forecasting, offering a scalable framework for near‐real‐time ozone prediction and supporting timely environmental and public health decision‐making.
Showing the abstract — retrieve the full paper via the Exa API.