Multi-granularity PM2.5 concentration long sequence prediction model combined with spatial–temporal graph

Abstract

Air pollution problem seriously affects the ecological environment and human health. More accurate predictions over a longer time span would enhance the effectiveness of early warning and prevention measures. Although existing methods have made progress in short sequence prediction, the predictions on long sequences remain challenges due to information loss. In this paper, we propose a spatial–temporal graph-based long sequence air pollutant prediction model. The proposed model first downsamples the time series into different granularities to capture the temporal features. Then, we use the vector production method to construct a spatial–temporal graph for each granularity which combines spatial information with temporal information. The unique spatial–temporal relationships of each city under different time granularities can be extracted by graph attention network (GAT). This approach helps model to capture dependencies in the time series comprehensively, thereby improving the accuracy of long sequence prediction. Based on the scenario and air pollution datasets imported from the detection station in Shanghai, extensive experiments show that the proposed model outperforms existing approaches on MSE and MAE.