Research On PM_2.5 Forecasting Based On Spatio-Temporal Data Mining

Air pollution has been a hot issue in recent,and PM_2.5has received extensive atten-tion as one of the main pollutants in the air.It is a great convenience for people to make timely decisions about production and life by knowing the future air quality in advance.Therefore,it is significant to make timely and accurate prediction of PM_2.5.Consider-ing the changing characteristics of PM_2.5in both temporal and spatial dimensions,PM_2.5prediction can be regarded as a spatio-temporal data mining task.However,the PM_2.5concentrations are affected by many factors such as meteorological conditions and geo-graphical environment,and the PM_2.5concentration in the target area may also originate from the spread of adjacent areas.The PM_2.5prediction problem still faces many chal-lenges.Existing PM_2.5prediction methods are mainly based on geographic information and other elements to build a fixed graph,but it is difficult to capture the dynamic corre-lation of PM_2.5concentrations between city nodes.Aiming at the above problems,this paper proposes two PM_2.5concentration prediction models based on the dynamic spatio-temporal graph network mainly from the perspective of the spatio-temporal correlation mining of PM_2.5concentrations between cities.The main work and innovations of this paper are summarized as follows:（1）This paper proposes a prediction model based on a two-stage dynamic graph network（DDGNet）.The model consists of two stages.In the first stage,we propose a method of using spatio-temporal weighted K-nearest neighbors,which can be measured from two dimensions of spatial and temporal to determine the most relevant neighbor nodes of the target node at any time to build a dynamic graph;in the second stage,we use the graph attention network to dynamically aggregate information from neighbor nodes.In addition,we also used a GRU-based recurrent neural network for modeling the tempo-ral dependences of PM_2.5concentration.（2）This paper further proposes a multi-graph fused dynamic spatio-temporal net-work（MDSTN）.This method not only takes into account the advantages of dynamic graphs,but also fully exploits the rich spatial relationship of city PM_2.5concentrations from multiple perspectives,and has stronger prediction performance.In terms of tempo-ral dependencies,we use gated temporal convolution to obtain the long-term dependence of PM_2.5concentration,which has better performance and faster speed than the RNN method;in terms of spatial correlations,we construct graphs from the perspectives of temporal,spatial and dynamic adaptation respectively,and propose a multi-graph convo-lution fusion module to fully mine the spatial correlations between nodes.In addition,we also fuse auxiliary features with PM_2.5features through a fusion mechanism to assist prediction.We compare the proposed model with related research work on a large real-world air quality dataset.The experimental results show that DDGNet can effectively capture the spatio-temporal dynamic correlation of PM_2.5between cities,MDSTN can fully exploit the diversity of spatial relationships of PM_2.5concentrations in cities,and the prediction results of our proposed models are better than existing methods under different prediction time periods.