Traffic Flow Prediction Based On Deep Learning

As an important supporting technology for urban intelligent transportation system to realize traffic guidance,traffic flow prediction is the key to alleviating traffic congestion and causing related problems.In recent years,the deep learning methods have gained more and more attention in traffic flow prediction due to its powerful nonlinear fitting and deep feature expression ability.Based on the unique spatial and temporal characteristics of traffic data,this paper attempts to construct more efficient input features to make research on the traffic flow prediction problem under the different data quantities that can be used,combining with the characteristics of different depth learning methods.The main work and research content are as follows:1.Crawl real-time traffic data on the website,then identify and repair missing and abnormal data to improve data quality.The spatial-temporal correlation of traffic flow is analyzed and summarized,not only correlation in the adjacent roads,but also has characteristics with temporal closeness,period and trend.2.When historical traffic data is limited and only temporal closeness of traffic flow is considered,the length of the input sequence has an important influence on the performance of the model.Most existing prediction models depend on experience method but lack of theoretical support when selecting temporal and spatial characters of traffic flow,this paper proposes a traffic flow prediction method based on spatial-temporal strongly correlation and long short term memory(LSTM)network.Combined with the prediction ability of different time lags and spatial(adjacent roads)characteristics of traffic flow,a spatial-temporal strongly correlative matrix is constructed to be the input of LSTM network in the form of time series to learn the temporal and spatial variation of traffic flow.Finally,experiments show that the method has lower prediction error and can effectively reduce the interference of redundant features.3.When historical traffic data is relatively sufficient,the spatial-temporal characteristics of traffic flow can be more fully utilized.In view of the existing methods,the temporal period and trend characteristics of traffic flow are less considered,and the long-term dependence oftraffic flow is difficult to solve.Using the idea of convolutional neural network(CNN)to recognize images for reference,a traffic flow prediction method based on multi-channels spatial-temporal convolutional neural network(MCST-CNN)is proposed.The temporal closeness,period and trend of traffic flow combined with spatial features are regarded as different channels input of the image respectively.And the deep CNN is designed to capture the global information in the input range,then learn the deep temporal and spatial dependence of traffic flow simultaneously.Finally,experiments show that the proposed method has higher prediction accuracy and stability.