Air Quality Prediction with A Meteorology-Guided Modality-Decoupled Spatio-Temporal Network

Air quality prediction plays a crucial role in public health and environmental protection. Accurate air quality prediction is a complex spatiotemporal problem, that involves pollutant spatiotemporal dependencies and meteorological influences that govern pollutant dispersion and transformation. However, existing predictive models often underestimate the critical influence of atmospheric conditions by treating air quality and meteorological data as a single, homogeneous modality or by using only limited surface-level features, which impairs performance. To overcome this, we propose MDSTNet, an encoder-decoder framework that takes historical air quality observations and multi-pressure-level meteorological data as input, explicitly modeling them as distinct modalities, while uniquely leveraging multi-step weather forecasts as dynamic prompts to guide prediction. Meantime, we construct ChinaAirNet, the first nationwide dataset combining air quality records with multi-pressure-level meteorological observations. Experimental results on ChinaAirNet and a public dataset KDDCUP-Beijing demonstrate MDSTNet’s superiority, substantially reducing 48-hour prediction errors by 17.54’ compared to the state-of-the-art model, showcasing the significant advantage of our meteorology-guided, modality-decoupled design. Code is available at this repository<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>, and dataset is available at here<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>.