Multi-Scale Graph Transformer Networks for Regional Emission Spatiotemporal Prediction

The spatiotemporal graph prediction of regional emissions offers scientific insights for designing sensible macro emission reduction policies and investigating regional emission characteristics’ spatiotemporal heterogeneity and dependency. First, the natural hierarchical structure of regional emissions is overlooked by existing methods., which can extract features effectively while being applied only to the micrograph of local networks. Second, emission monitoring sites’ widespread coverage and the accumulation of extensive dynamic data present challenges. The high-dimensional, nonlinear, and spatiotemporal correlations in emissions further complicate the capture of long-term spatiotemporal features or periodic patterns within graph data. To this end, we model the emissions as a microscale spatiotemporal graph sequence. And then obtaining macroscopic spatiotemporal graph sequence by using K-means clustering. Finally, both the micro and the macro data are used to extract long-term dependencies through Graph Transformer. Our method is validated on AQI emissions from the Guangdong province and outperforms state-of-the-art baselines on the spatiotemporal prediction method.