Research On Short-Term Traffic Flow Prediction Method Based On Deep Neural Networks

The application of intelligent transportation systems for scientific control of traffic flow is an important way to improve the efficiency of the road network.Timely and accurate short-term traffic flow prediction data can not only be used to improve the level of traffic signal control on the road network,but also provide a decision basis for traffic management departments to understand and grasp the traffic situation on the road network and make preventive traffic guidance in advance.Benefited from the rapid development of deep learning technology,short-term traffic flow prediction methods based on deep neural networks have become a hot research topic in recent years.However,when solving the traffic flow prediction problems of isolated single intersections and complex road networks,the prediction accuracy of the existing deep neural network-based short-term traffic flow prediction methods still needs further improvement.Therefore,this paper designs a new short-term traffic flow prediction algorithm,which digs deeply into the complex space-time correlation characteristics between different phases in isolated single intersection and different traffic nodes in the road network to accurately predict the short-term traffic flow information of the target object.The main research contents include:(1)A review of traffic flow prediction methods and comparative simulation analysis.This paper presents a review of common methods used in short-term traffic flow prediction research,including traditional short-term traffic flow prediction algorithms and deep neural network-based short-term traffic flow prediction algorithms.The model principles and application methods of Long Short-Term Memory(LSTM),Gated Recurrent Units(GRU),Sequence to Sequence(Seq2Seq)and Graph Convolutional Neural Networks(GCN)are elaborated.In order to verify the effectiveness of deep neural network models for short-term traffic flow prediction,simulation and analysis experiments are conducted on several publicly available datasets in this paper.(2)Short-term traffic flow prediction at single intersections.Existing short-term traffic flow prediction methods pay less attention to the phase flow relationships within isolated intersections,and it is still a challenge to further improve the prediction accuracy of short-term traffic flow at isolated single intersections due to the large random fluctuation and weak correlation of each lane flow within isolated single intersections.To address the above challenges,a multi-task parallel learning short-term traffic flow prediction method(MTL-fusion)based on LSTM and Seq2 Seq models is proposed in this paper.In the traffic signal control,the phase traffic flow is equal to the sum of the traffic flow of its internal lanes.Unlike the traditional traffic flow prediction methods that directly use the traffic flow of each lane as the minimum prediction unit,the MTL-fusion method proposed in this paper uses the traffic flow of each phase as the minimum prediction unit to solve the problem that the prediction accuracy is easily caused by excessive fluctuations of traffic flow when predicting the traffic flow of a single lane.The problem of low prediction accuracy due to the fluctuation of traffic flow is solved.Finally,the proposed traffic flow prediction algorithm is simulated and tested using the real traffic flow data collected from the intersection of Changsheng South Road-Huayan Road in Jiaxing City.The test results show that compared with the Seq2 Seq model-based prediction algorithm,the proposed MTL-fusion method can achieve better prediction results at different sampling time steps,and the evaluation indexes MSE and MAE are reduced by 7.88% and 15.54%,respectively.(3)Short-term traffic flow prediction of road networks.In the traffic flow prediction of road networks,the accuracy of the traffic flow prediction algorithm is directly affected by the ability to accurately model the spatial association relationships among the traffic nodes in the road networks.However,in the existing GCN-based short-term traffic flow forecasting methods,the GCN predefined parameter matrix can only model the explicit association relationships among the directly physically related nodes in the road network,and does not consider the hidden association relationships among the non-directly physically related traffic nodes in the road network.In this paper,we propose a short-term traffic flow prediction method(MGCN-fusion)based on a multi-graph convolutional neural network fusion model.The method models both explicit and hidden correlations between nodes in the road network by constructing Multi-graph Convolutional Networks(MGCN),and uses GRU to model the temporal correlations between traffic data of different traffic nodes in the road network.In addition,a fusion strategy with auxiliary loss is designed to achieve the fusion of the prediction results of each module within MGCN-fusion.Finally,simulation tests are carried out on typical public datasets PEMS04 and PEMS08 to verify the effectiveness of the proposed algorithm.Compared with LSTM,the MGCN-fusion model reduces the three metrics(MAE,MAPE,and RMSE)by 4.93%,15.26%,and 4.55%,respectively,on the dataset PEMS08.