Research On Spatial Temporal Prediction Model Of Traffic Flow Based On Attentional Mechanism

The advancement of Internet of vehicles technology promotes the development of intelligent transportation system(ITS).Therefore,the prediction model based on intelligent algorithm has become a research hotspot in ITS field.However,the spatial temporal dynamics of traffic flow data brings challenges to the design of prediction model.The non-spatial sensitive prediction model is unable to learn the dynamic influence between different locations in the actual traffic scene,and the existing spatial temporal prediction model does not make full use of the spatial temporal dynamic characteristics,so it is difficult to mine the spatial temporal model of traffic flow data.Therefore,it is of great research value to construct an appropriate spatial temporal model for traffic flow prediction.In this paper,the spatial temporal model is used as the center to study traffic flow prediction.Specific work and contributions are as follows:1)Through quantitative and qualitative analysis of the time series periodicity of traffic flow data,this paper shows that the daily periodicity and the weekly periodicity have similar data patterns,which can be combined into a periodic component;Second,cyclical components can be viewed as "future" data for recent components.Based on the analysis results,two modeling methods are proposed to solve the problem of periodic modeling of prediction model.Firstly,the existing multi-component modeling method is simplified to the two-component modeling method,and then to the single-component modeling method,so as to effectively improve the computational efficiency of the model and save the prediction time.2)Proposed an improvement graph spatial temporal convolutional neural network(GSTCNN).Firstly,an edge-based subgraph sampling method,EBSGS,is defined to improve the scalability of spatial model.Then a simplified topology adaptive network(STAGCN)is used to extract the spatial features of the subgraphs.Finally,the interpretable time series convolution network(TCN)is introduced into extract the time series features.3)On the basis of GSTCNN,proposed an improvement graph spatial temporal convolution network AGSTCN based on attention mechanism is proposed.Based on self-attention,a spatialtemporal separation attention mechanism is designed.Temporal attention is used to capture the dynamic dependence of the temporal dimension of traffic flow,and spatial attention is used to learn the dynamic influence weights between different locations in the traffic network.The spatial temporal attention mechanism in this paper can not only capture the dynamic influence of different spatiotemporal positions on the current spatiotemporal position,but also improve the remote dependency ability of GSTCNN.4)Experimental verification was carried out on two publicly available expressway traffic flow data sets Pe MSD8 and Pe MSD4.The experimental results show that: firstly,the average prediction time(MPTE)evaluation index performance of the coordinate axis and fusion input periodic modeling method proposed in this paper is superior to other methods on the two data sets.Secondly,the GSTCNN model improves the scalability of the model when the prediction performance is comparable.Finally,the prediction effect of AGSTCN model on two data sets was better than that of other models.The MAPE evaluation index scores on Pe MSD8 and PEMSD4 data sets were 11.05 and 13.76,respectively,which were 4.0% and 2.5% higher than the existing ASTGCN model.