Multi-task Air quality Network Prediction Based on Dynamic Spatio-temporal Graph Convolution

Abstract

With the rapid advancement of industrialization and urbanization, the concentrations of PM2.5 and PM10 have increased significantly, posing severe threats to both public health and economic development. PM2.5 and PM10 concentration levels serve as critical indicators of air quality, and accurate prediction of these pollutants is essential for effective mitigation strategies. Air quality is a dynamic parameter that fluctuates across both time and space, and traditional models often fail to capture its complex spatiotemporal relationships. To address this issue, this paper proposes a dynamic spatiotemporal graph convolution-based multi-task air quality prediction network (ST-MAP Net). By incorporating data such as the geographic locations of monitoring stations, wind direction, and wind speed, the model constructs dynamic graph structures that better represent the temporal and spatial variability of meteorological conditions. A time-space-time enhanced spatiotemporal convolutional block is designed, and a dynamic gating mechanism is introduced to effectively capture temporal features. Additionally, the model leverages a multi-task learning framework to facilitate joint optimization across multiple tasks. The ST-MAP Net is used to simultaneously predict the PM2.5 category for the upcoming hour, as well as the concentrations of PM2.5 and PM10. Experimental results demonstrate that ST-MAP Net outperforms existing models in multi-task forecasting, achieving a classification accuracy of 97.82%, an R2 of 0.9664 for PM2.5 concentration, and an R2 of 0.9578 for PM10 concentration. Ablation studies further confirm the efficacy, flexibility, and promising application potential of dynamic graph convolution and multi-task prediction mechanisms.