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

TL;DR

MDSTNet addresses the problem of air quality prediction by explicitly decoupling air pollutants and meteorological data into distinct modalities and leveraging multi-step weather forecasts as dynamic prompts. It introduces a three-branch encoder with Distillation Attention Modules to model spatial, pollutant-variate, and meteorological correlations, followed by a decoder that conditions predictions on forecast information. The authors present ChinaAirNet, a nationwide dataset that combines air quality records with multi-pressure-level meteorology, enabling large-scale evaluation and reproducibility. Experimental results show a 17.54% reduction in 48-hour prediction error over the prior state-of-the-art on ChinaAirNet, with ablations confirming the value of multi-level meteorology, cross-modal attention, and forecast-guided prompts for accurate, horizon-aware predictions.

Abstract

Air quality prediction plays a crucial role in public health and environmental protection. Accurate air quality prediction is a complex multivariate spatiotemporal problem, that involves interactions across temporal patterns, pollutant correlations, spatial station dependencies, and particularly meteorological influences that govern pollutant dispersion and chemical transformations. Existing works underestimate the critical role of atmospheric conditions in air quality prediction and neglect comprehensive meteorological data utilization, thereby impairing the modeling of dynamic interdependencies between air quality and meteorological data. To overcome this, we propose MDSTNet, an encoder-decoder framework that explicitly models air quality observations and atmospheric conditions as distinct modalities, integrating multi-pressure-level meteorological data and weather forecasts to capture atmosphere-pollution dependencies for prediction. Meantime, we construct ChinaAirNet, the first nationwide dataset combining air quality records with multi-pressure-level meteorological observations. Experimental results on ChinaAirNet demonstrate MDSTNet's superiority, substantially reducing 48-hour prediction errors by 17.54\% compared to the state-of-the-art model. The source code and dataset will be available on github.

Figures (8)

• Figure 1: Architectural Overview of MDSTNet. The framework employs an encoder-decoder structure, with each encoder layer comprising three parallel branches: Spatial Attention (Spa), Pollutant Variate Attention (Pva), and Meteorological Variate Attention (Mva). Each branch is implemented as an independent Distillation Attention Module (DAM). The decoder is also composed of a multi-layer transformer structure. The structure of DAM is illustrated in the upper right corner, while the lower right inset demonstrates the integration of Spatial and Variate Positional Encoding.
• Figure 2: Distribution of air quality monitoring stations in China and corresponding air quality levels.
• Figure 3: Temporal trends of AQI and pollutant concentrations at station A and station B.
• Figure 4: Specific humidity at multiple pressure levels over air quality monitoring stations.
• Figure 5: Effect of the Distillation Attention Module.