A PM2.5 spatiotemporal prediction model based on mixed graph convolutional GRU and self-attention network

The increase in atmospheric pollution has made it essential to develop accurate models for predicting pollutant concentrations. The current researches have faced challenges such as the neglect of significant information selection from local and neighboring stations, as well as insufficient attention to long-term historical data patterns. Therefore, this paper proposes a spatiotemporal prediction model called MGCGRU-SAN, which leverages long-term historical data to predict PM2.5 concentration values across multiple stations and multiple time steps in the future. Firstly, we employ the Mixed Graph Convolutional GRU(MGCGRU) module to capture the spatiotemporal dependencies in short-term historical time series from various stations. Secondly, the longterm PM2.5 historical time series (e.g. one week) is divided into uniformly sized segments and fed into the Self-Attention Network(SAN) module to capture the long-term potential temporal patterns. These enable the model to not only capture short-term fluctuations, but also identify and track long-term temporal patterns and trends in the prediction process. Finally, we conduct extensive comparative and ablation experiments using historical air pollutant and meteorological data from the Beijing-Tianjin-Hebei region. The experimental results demonstrate that the model, after capturing the long-term latent temporal patterns, achieve improvements of 9.62%, 6.33%, and 4.98% in the RSE, MAE, and RMSE evaluation metrics during multi-step prediction. Overall, the model outperforms the best baseline model by an average of 8.34%, 6.12%,4.06%, and 2.60% in RSE, MAE, RMSE, and Correlation metrics, respectively, showing superior performance in multi-station long-term predictions.