High-concentration ozone threatens human health and ecosystems, modulated by dynamic, multiscale meteorological processes. Existing machine learning studies for ozone prediction rarely incorporate the spatiotemporal evolution of regional meteorological fields (STRMFs), limiting the explanatory power of meteorological drivers in ozone variability. Thus, a sequential convolutional long short-term memory network framework (CNN-LSTM) was designed to utilize the STRMFs for ozone prediction. Scenarios incorporating STRMFs across multiple spatiotemporal scales were constructed using Global Forecast System (GFS) data sets. Model performance was evaluated in terms of ozone concentration prediction accuracy (AOCP) and precision in forecasting high-ozone pollution events (PHOE) across key Chinese regions. Appropriate expansion of meteorological data spatiotemporal scale enhanced AOCP, with notable improvements in PHOE, demonstrating ozone variability's dependence on multiscale meteorological processes. Leveraging meteorological data that better represent real atmospheric conditions improved AOCP. The CNN-LSTM framework explained over 85% of daily ozone variability through STRMF integration, successfully resolving how ozone concentration variations in key regions responded to typhoon positional shifts. This methodology enables timely pollution alerts while elucidating the critical role of regional meteorological processes in ozone pollution.