Traffic Forecast Based On Graph Convolution Network

Traffic prediction plays an important and fundamental role in an intelligent urban transportation system,and its core problem is to describe the complex temporal and spatial correlation of large traffic data.In recent years,compared with the traditional forecasting methods,the traffic prediction model based on Deep Learning model has significantly improved the prediction accuracy.However,the existing prediction models based on deep convolutional networks have difficulty in utilizing the spatial topology of road network segments and even destroy the spatial connection relationship of road network segments,so they are unable to effectively extract the local spatial features of road network segments and improve the accuracy of traffic prediction for road network segments.Considering the above problems,this paper focuses on a traffic prediction model based on a graph convolution network.The key research is on the graph structure representation of traffic network data,graph convolution network prediction model and multimodal data fusion.At last,this thesis achieves the following innovative results:First,a graph convolutional network traffic prediction model based on optimized and dynamic road network graph representation is proposed.Traditional traffic prediction models based on graph convolutional networks usually use a static graph representation based on the distance of road segments or the structure of the road network,ignoring the data correlation between adjacent road segments hidden in the sequence data of the road network.In response to the above problems,this dissertation proposes a graph convolution network traffic prediction model based on parametric road network maps and dynamic road network maps.By learning the correlation between the road network sequence data and combining the spatial characteristics of the original road segments,an optimized road network map is constructed,Hadamard product optimization road network map and dynamic road network map three road network map representations,so that the road network map structure can reflect not only the spatial distribution of road segments,but also the data distribution of road segments.Experiments show that the proposed road network graphical representation method works much better than the static road network graph representation method in the graph convolutional network traffic forecasting model.Second,a traffic prediction model based on the hierarchical road network graph convolutional network is proposed.The urban road network has significant regional transport characteristics.Due to the division of administrative regions,different regions naturally have different traffic characteristics and types.However,the traffic forecasting models based on graph convolutional networks usually model only the local traffic characteristics of road segments,while ignoring the macroscopic regional traffic characteristics mentioned above.Therefore,this dissertation proposes a traffic forecasting model based on a hierarchical road network graph convolutional network.The regional road network graph is obtained by segmenting and clustering the road network graph,and a hierarchical representation of the road network graph on two spatial scales is constructed.Then a hierarchical graph convolutional network traffic forecast model is built by transferring and linking the two spatio-temporal features of the area and the road segment,and finally a more accurate traffic forecast is realized.Third,an enhanced graph convolutional network traffic prediction model based on POI features is proposed.The urban transportation system is a complex system that is influenced by many factors,such as severe weather,major events,and the distribution of urban points of interest(POI)such as hospitals and schools,etc.However,most of the current traffic forecasting methods use only the observed traffic flow data and not the POI data mentioned above.The reason is that these POI data and traffic data have a heterogeneous gap and are difficult to use,and some traffic forecasting methods using these data have little improvement.Therefore,the use of cross-modal data such as POI becomes a difficult problem in traffic forecasting.In response to this problem,this paper proposes an associative learning network based on the data from POI and the state data of the road network,embeds the data from POI into the traffic prediction model of the graph network,and improves the feature expression ability of graph convolutional networks by data-driven adaptive methods.Tests on various traffic prediction models of graph convolutional networks show that the feature enhancement module POI proposed in this paper is a component for improving traffic prediction models of graph convolutional networks,considering both generality and effectiveness.In summary,this paper introduces the graph convolution model to describe the local spatial characteristics of road network data and conducts an in-depth study on the parameterized and dynamic graphical representation of road network,the regional hierarchical graphical representation of road network and the use of cross-modal road section POI data.Based on road network datasets of many cities at home and abroad,extensive verification experiments are conducted in this paper.The experimental results show that the graph convolution network based on the micro-dynamic representation of the road network and the macro-regional representation of the road network can greatly improve the accuracy of the traffic prediction model,and verify the effectiveness and universality of the cross-modal POI function expansion module in the graph convolution network prediction model.Finally,compared with the traditional prediction algorithm,this paper enables more accurate traffic prediction,which provides a solid foundation for building an intelligent transportation system.