Research On Urban Traffic Flow Forecast Method Based On GCN

Accurate prediction of traffic flow is a key part of the intelligent transportation system.It is important to analyze the traffic flow data with spatial and temporal characteristics in urban traffic system and extract the implied features from them to realize real-time scheduling of traffic conditions and urban road planning.With the large-scale deployment of road facilities,the amount of data is growing and the urban road conditions are becoming more and more complex.Although scholars at home and abroad have proposed many methods for traffic flow prediction,there are still many problems to be solved,including the difficulty of extracting spatial-temporal features,the lack of perfect extraction of road dynamic correlation features,and the defects of spatial feature extraction methods.Based on the graph convolutional neural network,this thesis focuses on spatial-temporal patterns,road topology and spatial-temporal prediction related techniques to dig deeper into the temporal,spatial and other hidden features in traffic data to achieve the improvement of prediction effect.The research in this paper focuses on urban road system,considering the characteristics of different temporal scenarios and the specific conditions of roads,and uses graph neural network for spatial-temporal prediction,which has reference value for similar temporal data research.The main research of this paper contains the following two aspects.1.To address the difficulty of capturing spatial-temporal dependence in traffic flow prediction problems,we design a multidimensional attention spatial-temporal network model named MA-STN.firstly,the spatial-temporal characteristics of traffic flow data exist are studied,and data sets of different dimensions are divided with reference to time periods.Secondly,the attention mechanism is used to obtain the weights of different nodes of the network and the weights on different time slices,respectively.Finally,the spatial dependency is extracted by combining the spatial weight matrix with the graph convolution GCN,and the temporal dependency is extracted by combining the temporal weight matrix with the Conv LSTM,and the features extracted from the data of different dimensions are fused to finally output the prediction results.Comparing with similar models,the proposed MA-STN model achieves better results on both real datasets.2.For the current graph convolutional neural network only captures the spatial dependence on the static graph structure and ignores the potential relationship between nodes.Meanwhile,the traditional frequency domain graph convolution contains only low-pass filter cannot extract features with heterogeneity on the graph.Therefore,we propose a traffic prediction model based on dynamic traffic graphs.The model uses node embedding to construct an adaptive adjacency matrix,and is applied to a frequency adaptive graph convolution neural network proposed.After experimental analysis,the adaptive adjacency matrix can respond to spatial features beyond the fixed graph structure,and both high and low frequency features help the learning of node representation,and the traffic prediction model based on dynamic traffic map proposed can also predict the traffic flow more accurately.