Short-term Traffic Flow Deep Learning Prediction Models And Their Parallel Training

Real-time and accurate short-term traffic flow prediction is the premise of traffic control and guidance,which is of great significance to the alleviation of traffic congestion and development of social economy.The emergence of deep learning provides new ideas for traffic flow prediction.In this paper,two networks of the deep belief network(DBN)and stacked autoencoder(SAE)are used to construct the traffic flow prediction models.Traffic flow features are extracted by unsupervised learning method.However,the training processes of the prediction models based on DBN and SAE are time-consuming and cannot meet the real-time requirement of traffic flow prediction application systems.This paper proposes a parallel training strategy based on data parallel mode.The feasibility of parallel training is analyzed,and the formulas of parallel training are derived,which compresses the training time of the two traffic flow prediction models and improves the real-time performance of traffic flow prediction applications.The concrete research contents are as follows.Firstly,the traffic flow of a single section of a single road segment is selected as the research object.The traffic flow prediction models are constructed based on two deep learning networks.The training and testing processes of the models are accomplished.The testing results are compared with that of the traditional models of backpropagation neural network(BP-NN)and support vector machine(SUM).The simulation results show that the prediction accuracies of the two models based on deep networks are similar and better than that of the traditional models,which validates the advantages of the deep learning models in traffic flow prediction.Secondly,correlation methods are used to optimize the prediction models.Rectified linear unit(ReLU)is chosen as the activation function to replace the sigmoid function for the problem of gradient vanishing widely existing in deep networks.Momentum method is added to the gradient descent process to prevent the training process from falling into local minimum.Dropout method is used to eliminate the over-fitting phenomenon in the training process,which improves the generalization ability of the models.Finally,in order to solve the time-consuming problem of training processes of the above two deep learning models,this paper proposes a parallel training method based on data parallel mode.Parallel feasibility analysises of BP algorithm and contrast divergence(CD)algorithm used in the training processes are carried out.The parallel calculation formulas of the two algorithms are derived and applied to all phases of the training process.The training data is equally divided into n parts and distributed to n computing nodes.One of those nodes is called the master node,and the remaining are called the slave nodes.The role of the slave node is to extract the traffic flow features using the local subdataset.In addition to the function of the slave node,the master node also synthesizes the calculation results of each slave node and broadcasts the parameters to each slave node.The simulation results show that with this parallel method,the training time of the prediction models is reduced while the prediction accuracy is guaranteed.