Deep Learning-based Models And Methods For Travel Time Prediction

In recent years,the increasing number of vehicles in the road network has ag-gravated the uncertainty of traffic conditions,resulting in new challenges in traf-fic prediction in Intelligent Transportation System.Deep learning based model has strong ability of dependency modeling and feature extraction,thus it is widely ap-plied in many fields.However,there are still some problems in the field of traffic prediction:(1)Long short-term memory network prediction model uses its memory to realize long-term dependence modeling,which is difficult to reflect the depen-dence relationship between each historical traffic data point and prediction informa-tion.(2)The traffic predicting model based on convolutional neural network models the dependence relationship of traffic prediction by convoluting the adjacent ele-ments in the matrix,while the local receptive fields cannot reflect the dependence relationship between historical traffic data points and prediction information.(3)In traditional convolutional neural networks,the dependence relationship between con-volution results and prediction information is realized through full connection layer.This method does not pay adaptively attention to the features in convolution results,so the feature utilization is insufficient.To solve the mentioned problems,this dis-sertation studies the modeling of the dependence relationship in traffic data for traffic prediction,based on deep learning methods including long short-term memory net-work,convolutional neural networks and attention mechanisms.The main contents and contributions of this dissertation are summarized as follows.(1)A long short-term memory-based model with attention mechanism is proposed for travel prediction.On the basis of validating the dependency relation-ship between historical traffic data points and prediction information,the depth of network is constructed based on attention mechanism,and the long-term dependency relationship in traffic data is modeled.The results of comparative experiments show that the model has better prediction effect than existing long short-term memory net-work and other baseline methods.Case studies show that attention mechanism can focus on historical traffic data or features adaptively and make effective use of data or features.(2)A new method convolutional neural network to defining the dependence between historical traffic data points and prediction information,i.e.a new lo-cal acceptance domain,is proposed.Based on the new local acceptance domain,a convolution neural network-based model with single link for travel prediction is proposed.The model uses the new local acceptance domain in the time dimension to define the dependence between historical traffic data points and prediction informa-tion.The comparative experiments between the proposed model and the traditional convolutional neural network and other baseline methods show that the proposed model achieves better prediction accuracy.The effectiveness of the new local recep-tive field is approved by comparative experiments based on different links between the proposed model and the traditional convolution neural network on a certain num-ber of links.Further experiments show that the new local acceptance domain can reflect the dependence between each historical data point and prediction information in the input sequence.A convolutional neural network-based model with multiple links for travel time prediction is proposed.On the basis of verifying the dependence relationship be-tween historical traffic data points and prediction information on multiple links,the new local receptive fields are generated base on traffic data on multiple links.The features of traffic data are extracted by using convolutional neural network on the new local receptive fields.The comparative experiments between the proposed model,traditional convolution neural network,prediction model on single link and other baseline methods show that the proposed model achieves better prediction ac-curacy.(3)A convolution component-based model with attention mechanism for travel time prediction is proposed.The proposed model use attention mechanism on the convolutional result generated by convolution component,implying adaptive attention to convolution results,constructing the dependence relationship between the convolutional result and prediction information,when interval times as the in-put of the attention mechanism.The comparative experiment results show that the proposed model achieves better prediction accuracy than the baseline methods on 47 randomly selected sections.