Dual-channel spatial-temporal difference graph neural network for PM2.5 forecasting

Accurate PM2.5 forecasting is significant for improving quality of life and human health. However, it is very challenging to capture the high spatiotemporal correlations and the complex diffusion processes of PM2.5. Most existing PM2.5 prediction methods only focus on spatiotemporal dependencies. In addition, the PM2.5 diffusion process with domain knowledge in deep learning is rarely considered. Therefore, how to simultaneously capture comprehensive spatiotemporal dependencies and model the complicated diffusion process of PM2.5 is still a challenge. To address this problem, we propose a dual-channel spatial-temporal difference graph neural network (DC-STDGN) to forecast future PM2.5 concentrations. DC-STDGN first constructs a dual-channel structure to obtain distance-based local neighboring information and the global hidden spatial correlation of the data. Then, a temporal convolution layer is designed to handle the long-term dependency. Finally, the spatial difference with domain knowledge is introduced to model the complex diffusion process and capture more comprehensive spatiotemporal correlations. The extensive experiments with three real-world datasets demonstrate the improved prediction performance of DC-STDGN over state-of-the-art baselines. DC-STDGN outperforms the second-best model by up to 16.9% improvement in mean absolute error, 8.9% improvement in root mean square error and 18.2% improvement in mean absolute scaled error.