Research On Traffic Flow Prediction Method Based On Convolutional Neural Network And Attention Mechanism

As urbanization continues to advance,the ownership of motor vehicles is increasing while putting more pressure on the road system and causing many problems,such as environmental pollution,traffic safety,traffic congestion,etc.To effectively address traffic congestion,intelligent transportation systems(ITS)have been widely applied in dynamic traffic management,and traffic flow prediction is an indispensable and important component of these systems.Accurate traffic flow prediction can assist urban traffic management departments in traffic guidance and control,optimize road resource allocation,improve road traffic capacity,and effectively solve urban traffic congestion and traffic safety issues,reduce air pollution,and promote sustainable economic development.Therefore,in-depth research on traffic flow prediction is of great theoretical and practical significance.Most current traffic flow prediction methods do not fully consider the spatio-temporal characteristics of traffic flow,and some prediction methods only consider the static spatial correlation of traffic road networks,but they do not the dynamic spatio-temporal correlation of traffic flow.Therefore,this study aims to apply deep learning methods to traffic flow prediction in order to effectively capture the spatio-temporal characteristics of traffic flow and thus improve the prediction accuracy and practicality.The main research of this thesis is as follows:(1)To address the problem of insufficient extraction of spatio-temporal features of traffic flow,this thesis proposes a combined traffic flow prediction model(TCMAN)based on temporal convolutional multi-attentive networks to effectively capture the spatio-temporal features of traffic flow.The model integrates a temporal convolutional network(TCN)and a codec-decoder structure(Encoder-Decoder),aiming to simulate the effects of spatio-temporal factors on traffic conditions.In the model,the TCN is used to capture the temporal features of traffic flow,while the codec structure is used to capture the spatial features of traffic flow and convert the encoded traffic features into future time-step sequence representation as the input to the decoder through a transformed attention mechanism.In this thesis,extensive experiments are conducted on METR-LA and PEMS datasets,and the results show that TCMAN has better prediction performance.(2)To address the problem that the dynamic spatio-temporal correlation of traffic flow is difficult to extract,this thesis proposes a traffic flow prediction model based on a timeconvolutional graph ODE network called G-ode Wave Net.This model is able to capture the spatio-temporal correlation of traffic flow simultaneously by combining a time-convolutional network(TCN)with an ordinary differential equation(ODE).G-ode Wave Net learns the hidden spatial features from traffic flow effectively by building an adaptive adjacency matrix to learn the hidden spatial relationships,effectively learning the hidden spatial features from the traffic flow.The prediction performance of the G-ode Wave Net model is validated on the METR-LA dataset and the PEMS-BAY dataset,and the experimental results show that the G-ode Wave Net model has high accuracy in traffic flow prediction.In summary,this thesis investigates the traffic flow prediction problem based on deep learning,making full use of its powerful nonlinear fitting ability and deep feature representation of data.Meanwhile,it is proved through experiments that the traffic flow prediction method based on convolutional neural network and attention mechanism proposed in this thesis can effectively predict the traffic flow.