An efficient modern convolution-based dynamic spatiotemporal deep learning architecture for ozone prediction

Abstract

Ozone pollution threatens ecosystems and human health, necessitating accurate forecasting for better management and policy implementation. To address this, we developed O3ConvNet, a convolution-based dynamic spatiotemporal deep learning model. It incorporates ModernTCN, a multivariate time series feature module, and a spatial message passing module using a dynamic adjacency matrix with geographic and DTW-based distances. O3ConvNet balances performance and efficiency across datasets with varying station densities and data qualities. In Los Angeles, the mean absolute error ranges from 6.984 μg/m³ to 15.990 μg/m³ for 1-h to 24-h predictions, with R² values exceeding 0.937. Computational time is reduced by up to 82% compared to the best baseline model. In Wuxi, China, it improves prediction accuracy by 18% and efficiency by 81%. ModernTCN module identifies critical factors for ozone formation, while the dynamic adjacency matrix helps extract spatial dependencies effectively. Overall, this study introduces a robust and generalizable model for regional ozone predictions.