Study On Expressway Export Flow Prediction Based On Deep Learning Of Spatio-temporal Characteristics

Expressway congestion is a traffic problem faced by the whole country,which has become national,normal and serious.Accurate traffic flow prediction can provide a basis for traffic management departments to take advanced measures and provide scientific guidance for effectively reducing traffic pressure.At present,Intelligent Traffic Management System plays an increasingly important role in solving traffic problems.As the foundation of Intelligent Traffic Management System--Traffic Information Collecting System,it collects traffic data effectively through scientific and technological means and equipment,and provides real-time and high quality data base for traffic forecasting.Toll station is the bottleneck node in the expressway traffic network,which is easy to cause the traffic queuing problem.The data collected from toll station can be used to predict the traffic volume,which can effectively alleviate the congestion problem at the export of toll station and provide convenient and efficient travel services for citizens.Based on the highway toll station data of one certain province,this paper establishes a toll station export flow prediction model by using deep learning method(Back Propagation Neural Network,Recurrent Neural Network and Long Short-Term Memory networks).K-fold cross validation and grid search algorithm are introduced in model training to obtain the best parameter combination.The specific prediction process of the model is as follows :Firstly,on the basis of analyzing the traffic space-time characteristics of the expressway toll station,the advanced flow of the correlated imports and the historical flow of the export are selected as the influencing factors for the prediction of the export flow.Secondly,the travel time is taken as the time surplus of the import and export flow sequence,and the improved Spatio-Temporal correlation coefficient is used to calculate the correlation coefficient of the associated imports flow sequence and export flow sequence after considering the time surplus,and to screen out the imports that are strongly correlated with export.Finally,the screened correlated imports flow and export historical flow are taken as the input of the neural network to predict the export flow.The correlated imports flows are regarded as spatial factors,and the historical export flows are regarded as time factors.Through the simulation analysis,it is concluded that :(1)If the Spatio-Temporal factor is taken as the input of three neural network models at the same time,the prediction results of LSTM and RNN are more accurate than BP neural network,but the prediction results of LSTM and RNN are very close.In the forecast period,the forecast effect of the weeks is better than that of the weekend.(2)Based on LSTM itself,time factor,space factor and Spatio-Temporal factor are taken as the input of the model respectively.The prediction results show that the prediction accuracy is the highest when considering Spatio-Temporal factor,the lowest when considering time factor,and the prediction accuracy is close when considering Spatio-Temporal factor and space factor respectively.(3)On the premise of considering the Spatio-Temporal factor,1 to 12 correlated imports were taken as the input of the model,and it was found that the more the number of correlated imports,the higher the prediction accuracy;(4)On the premise of considering the Spatio-Temporal factor,the flow from 5 minutes to 4 hours of the export history was taken as the input of the model,and it was found that the historical time window size of the export flow had no significant influence on the prediction results.(5)Divide 24 h of a day into 24 time features and take them as the input of the model together with the input vector of space-time.It is found that the time period features have an impact on the prediction results,but the effect is not obvious.