| With the improvement of people’s living standards and changes in travel patterns,the number of motor vehicles in our country has increased rapidly,which has given rise to a series of traffic problems.As an important measure that can alleviate traffic congestion at a lower cost,intelligent transportation has attracted more and more attention from researchers.Traffic flow prediction provides an important reference for traffic management and control in intelligent transportation by estimating the change trend of traffic flow in the road network,thereby improving the effective utilization of the road network and reducing traffic congestion in the road network.By studying the spatiotemporal relationship of traffic flow data,this paper proposes an efficient longrange prediction model of traffic flow based on fusion correlation,which improves the prediction performance and prediction speed of the model,and provides decision support for accurate and efficient traffic control.The main research contents and results are as follows:(1)To address the problem that simple adjacency relations cannot finely model complex road networks,a traffic flow prediction model Fusion-ST based on fusion correlation is proposed.The model is mainly composed of graph convolutional neural network and self-attention.In order to improve the accuracy of traffic flow prediction,this paper analyzes the spatiotemporal characteristics of traffic flow data using correlation functions from various aspects.Based on the analysis of the time relationship of hours,days and weeks in the traffic flow data,this paper focuses on the systematic analysis of the spatial relationship,including: the k-order neighborhoods of road network nodes,the shortest path of nodes and the similar regions of nodes,etc.,and effectively fuse the correlations among them.The local spatial relationship of nodes is obtained by using korder neighborhood,the topology of the road network is obtained by using distance matrix,and the global spatial relationship in the road network is obtained by using regional similarity relationship,which guarantees the accuracy of road network modeling by fully considering the complexity of spatiotemporal relationship.Finally,validation was performed on publicly available traffic flow datasets,and the results showed that the Fusion-ST proposed in this paper outperforms six baseline models in MSE,RMSE,and MAPE metrics,18.41,30.78,and 12.44,respectively.(2)Based on the model Fusion-ST,this paper proposes methods for the two modules of data input and prediction methods.Aiming at the problem that traffic flow prediction cannot solve the periodic time relationship in data,a multi-scale traffic flow prediction model Multi-Fusion-ST based on empirically divided data is proposed on the basis of model Fusion-ST.The improved multi-scale module is used in data selection,and traffic flow sequences of different periods are input into the model to extract temporal features,which improves the model’s ability to obtain periodic temporal relationships.To address the problems of slow prediction speed and the accumulation of errors in long-range prediction of the model Fusion-ST,a mask-based efficient traffic flow long-range prediction model Fast-ST is proposed to achieve the input data denoising while guiding the model to learn the features of the data by using a combination of dynamic learning rate and dynamic mask rate each other.Finally,it is verified on the public data set,and the results show that the Multi-fusion-ST modeled using the multi-scale characteristics of the data improves 1.2%,1.05%,and 0.3% in MSE,RMSE,and MAPE metrics,respectively.The mask-based traffic flow prediction model Fast-ST improves the prediction speed by 3-6 times under the condition of ensuring the prediction accuracy. |
PDF Full Text Request