Research On Traffic Flow Prediction Of Urban Road Networks Based On Capsule Networks

The dramatic increase in motor vehicle ownership has led to problems such as congestion in urban traffic,especially in urban traffic areas with complex road network structures,and the research and application of intelligent transport systems has become an important tool in dealing with urban traffic problems.Real-time and effective traffic forecasting can not only provide traffic management with current and future traffic conditions on road secti ons,but can also provide travel route selection and planning for travellers,alleviating urban road congestion,which is also of realistic social significance in building smart cities.At present,the complexity of urban traffic road networks is increasin g,and for road networks with hundreds or thousands of links in the region,how to efficiently explore the deeper hidden spatio-temporal correlation information in traffic flow data and build a regional traffic predictio n model has become an important task in traffic flow prediction.Based on this,this paper takes the urban regional traffic road network as the research object,fully learns the spatio-temporal information of traffic flow between each link road of the regional complex road network,construct s a prediction method based on Caps Net(Capsule Network),and integrates techniques such as attention mechanism and long and short term memory network(LSTM)to study the traffic flow of the link structure of the urban r oad network.The main work is as follows:(1)Based on the time series data,this paper converts the traffic flow data into images and processes the road network node relationships and spatial structures contained in the images based on Caps Net’s good spat ial location sensing capability.For the two datasets with different road network structures,this paper adopts different image conversion methods to convert traffic speed data into images,and Caps Net is used to process the spatio-temporal information of the traffic it contains and extract various features between road network links.Experiments are also conducted on the two datasets to analyse the predictive performance of Caps Net in traffic flow.The experimental results show that Caps Net has superior pe rformance for feature extraction of two different types of road netw ork structures.(2)To address the problem of predicting traffic flow on urban road networks at a regional scale,this paper proposes a regional traffic speed prediction method that combines Caps Net and an attention mechanism.Based on the image transformation method,Caps Net is used to analyse image data containing spatio-temporal information,learn and capture the spatio-temporal features of traffic flow in a specific region,and embed the attention mechanism into Caps Net t o aggregate local features of traffic and enhance the extraction of local features by the underlying convolutional layer of Caps Net to achieve the prediction of regional traffic speed.The experimental results show that the proposed model has the ability t o characterise a larger range of traffic features.(3)On the basis of general road network prediction in urban areas,the traffic prediction scenario is expanded and a combined prediction model of Caps Net fused with deep bi-directional LSTM(D-Bi LSTM)is proposed for traffic complex road networks.By focusing Caps Net on the learning of traffic spatial features and considering the backward dependence of time series data,the Bi LSTM is used to capture the temporal correlation information between links and introduce an overlay mechanism to build a deeper prediction model,which can fully handle the feature information of complex link structures and improve the prediction capability of the model.The experimental results show that the method has good prediction performance for traffic flow prediction of complex r oad networks,and the prediction accuracy is higher than that of the benchmark model.