| The acceleration of urbanization and the growth of per capita vehicle ownership make comprehensive urban management face enormous pressure in modern society.Traffic management is one of the essential tasks of urban management.Its goal is to reduce the economic and human losses caused by traffic congestion,exhaust emissions,and traffic accidents in cities.With the rapid development of computer and Internet of Things technologies,intelligent transportation systems play an increasingly important role in urban traffic management and smart city construction.Traffic flow prediction is the foundation of intelligent transportation systems,and accurate traffic prediction is essential for many applications.At present,traffic forecasting is still very challenging,mainly in two aspects: First,due to the complex spatiotemporal correlation and nonlinear changes in the traffic network,the prediction accuracy of existing models is not ideal,especially in long-term traffic forecasting.Second,the current leading research focuses on improving the accuracy of traffic prediction models while ignoring the time constraints of traffic prediction in specific scenarios to a certain extent.Aiming at the above problems,this paper proposes three new traffic flow prediction algorithms and validates the superiority of the proposed algorithms through experiments on realworld datasets.The main work is summarized as follows:(1)Aiming at the two main problems in the current traffic flow prediction algorithms:insufficient modeling of spatiotemporal correlation and low efficiency of learning long-term temporal dependence,this paper proposes an asynchronous dilated spatial-temporal graph convolutional network(ADGCN).ADGCN uniformly expresses spatial-temporal correlations in transportation networks as asynchronous spatiotemporal connections and proposes asynchronous spatial-temporal graph convolutions to model them accurately and flexibly.Subsequently,to solve the problem of over-fitting and improve the long-term prediction ability of the ADGCN model,this paper generalizes the one-dimensional dilated causal convolution to the form of graph convolution and proposes asynchronous spatial-temporal dilated causal convolution to utilize the shallower network to obtain a larger receptive field.Experimental results on three real-world traffic datasets show that ADGCN outperforms comparable models.(2)To solve the problems that existing traffic flow prediction algorithms measure the correlation between traffic nodes in a single way and ignore the rich semantic information between nodes,this paper proposes a traffic flow prediction model based on multi-view graph convolution and attention mechanism(TGANet).Multi-view graph convolution uses multiple metrics to evaluate the association strength between traffic nodes to construct an adjacency matrix of the traffic network.Furthermore,it combines the multi-head attention mechanism to model the global correlation,local correlation,and fine-grained correlation,respectively.Furthermore,to reduce the high computational complexity of the attention mechanism,a mask operation is added for multi-head attention.The mask operation filters out irrelevant information and speeds up the computation of the attention mechanism.Experimental results show that TGANet outperforms other baseline models in most cases.(3)To settle the problem that most traffic flow prediction models have too much time overhead in specific scenarios,this paper proposes a deep learning model FedAGCN based on federated learning and asynchronous graph convolutional networks.FedAGCN divides the traffic network into multiple sub-graphs.Each sub-graph is regarded as an independent traffic graph to train the submodel and the cloud model,which significantly reduces the model’s time overhead and deployment cost.Furthermore,to take advantage of the topology of the traffic network,a graph federated learning strategy,Graph Fed,is proposed to solve the parameter update problem between sub-graphs.At the same time,FedAGCN improves the ADGCN model in Chapter 3 and uses the data-driven node similarity measure method to construct the adjacency matrix and asynchronous spatial-temporal relationship matrix of sub-graphs.Experiments show that,FedAGCN greatly reduces the model’s training time and inference time while ensuring higher prediction accuracy. |
