Road Multi-period Traffic Flow Prediction Based On Deep Spatiotemporal Neural Network

In recent years,with the continuous increase in the number of motor vehicles,road traffic congestion has become increasingly serious.Scientific and accurate prediction of road traffic flow can timely solve road traffic congestion and play an important role in path planning and other aspects.Traffic flow prediction is also an important research field for smart transportation and smart cities.With the development and progress of society,the structure of urban road networks is becoming increasingly complex,and the field of traffic flow prediction is also facing huge challenges.In terms of time dimension,road traffic flow undergoes dynamic changes over time and has a certain periodic pattern.In terms of spatial dimension,simply considering the spatial relationships between upstream and downstream segments of predicted road sections can no longer meet the complex road network relationships today.Traditional methods of obtaining road network spatial information features through Convolutional Neural Networks(CNN)can no longer meet the requirements.In order to adapt to the increasingly complex road network relationships and improve the accuracy of traffic flow prediction models,this paper models the road network as a graph structure and designs a new spatiotemporal neural network traffic flow prediction model that integrates multi-source information based on the Graph Convolutional Neural Networks(GCN).First,on the data definition,the distance factor is integrated to design the adjacency matrix of the road network diagram.Secondly,in the construction of the network model,a spatial graph convolution module and a time series prediction module are designed to extract the spatiotemporal characteristics of traffic flow,respectively.Finally,improve the algorithm structure of the gating mechanism to enable the model to simultaneously receive input from multiple sources of information.The main research content of this article is as follows:(1)In terms of data definition,in the construction process of road network structure,traditional methods only focus on whether the nodes of the road section are adjacent,ignoring the differences in spatial influence between adjacent nodes.Therefore,by considering the influence of distance factors between adjacent nodes on spatial correlation,this paper designs a differentiated graph adjacency matrix integrating distance factors to store spatial road network information in a refined way.(2)In terms of network model construction,in order to more effectively explore the spatiotemporal characteristics of traffic flow data,spatial graph convolution modules and time series prediction modules are designed separately.The spatial graph convolution module is designed based on the graph convolution neural network.The spatial feature information is extracted through the graph adjacency matrix of the fusion distance factor;The time series prediction model is designed based on Gated Recurrent Unit(GRU)to extract feature information from the time dimension of data traffic flow.(3)In order to enable the network model to simultaneously input multi-source information and explore more logical relationships between data,this paper improves the algorithm structure of the GRU network unit gating mechanism to achieve the goal of multisource information fusion input.Firstly,it is necessary to receive the spatial feature information of the road network obtained by the spatial graph convolution module;The second is to consider the periodic characteristics of traffic flow and introduce periodic sequence information of traffic flow;The third is to receive feature information transmitted by the previous hidden layer unit in the network layer.This article conducts experimental verification on a real traffic flow PEMS dataset,in order to comprehensively evaluate the prediction accuracy of the model,and therefore verify the error performance of the model at different prediction periods.And select other graph convolutional network models proposed in the field of traffic flow prediction as comparative models,and design relevant comparative experiments.The experimental results show that compared to the optimal error results of other comparative models,in the test results of multiple datasets,the RMSE index of the model designed in this paper decreased by 2.3%-3.5% and the MAE index decreased by 1.2%-3.7% in the short period of 15 minutes.In the5-60 minute multi period result measurement,the RMSE error index decreased by 0.9%-1.4%,and the MAE index decreased by 0.7%-1.3%,further reducing the prediction error of the model.