PM2.5 prediction based on dynamic spatiotemporal graph neural network

Abstract

Air pollution is one of the main public health and safety issues facing humanity. PM2.5 concentration prediction (PCP) helps the public to prevent and make government decisions in advance. PCP is a typical knowledge mining problem based on spatiotemporal sequential data, which still faces great challenges up to now. Aiming at the complex conundrum of meteorological, geographical, and temporal factors interference and concentration sudden changes, a dynamic spatiotemporal graph neural network (DST_GNN) method for PCP is proposed by using the advantages of graph neural network (GNN) and mechanism model. Its main methods are: The graph structure is used to construct the spatial relationship of PM2.5 among different monitoring stations, the mechanism model HYSPLIT is used to construct the dynamic edge relationship among graph nodes, and the gate recurrent unit of attention mechanism is used to learn the timing of PM2.5 concentration, thus forming a GNN architecture that integrates machine learning and domain knowledge. In addition, a loss function based on trend and shape is proposed when the model objective function is designed. The proposed model innovatively uses HYSPLIT to assist in building a dynamic spatiotemporal graph network and uses trend loss function for model training, which provides a new way for the dynamic construction of GNN, and provides a reference for PCP by combining domain knowledge and deep learning. Experimental results show that the proposed method has the best prediction accuracy among GNN based methods, which reduced the mean absolute error by about 14% and root mean square error by about 13% compared with the advanced GNN methods. The mean absolute error within 48 h forecast is less than 50, which predictive performance is far superior to the traditional mechanism model, and it also has the characteristics of flexible deployment and easy implementation.