Research On Forecasting Methods Of Short-term Traffic Flow Based On Temporal-spatial Information

Motor vehicles have been widely used with the steady development of economy and technology in recent years.However,the increased use of vehicles creates serious issues,including vehicle congestion and accidents.Intelligent Transportation System(ITS)has become an effective way to solve traffic congestion and improve the quality of traffic management by virtue of advanced artificial intelligence and computer communication technology.As an important function of its application,short-term traffic flow prediction can provide accurate and timely traffic state information that the system relies on.Based on the above discussion,the short-term traffic flow prediction models based on spatiotemporal information are deeply studied in this paper,and the specific work is as follows:Based on the theory of short-term traffic flow prediction,the basic characteristics,parameters and prediction process of short-term traffic flow are described in detail.And the classical prediction models such as ARIMA,SVR and SAEs are introduced,which are convenient for the subsequent experimental comparison.At the same time,two kinds of highway traffic data sets provided by TDRL and Pe MS are introduced,as well as the data preprocessing ways to ensure the integrity of the data set and the performance evaluation index used to evaluate the prediction effect of models,which lay the foundation for the later research work of short-term traffic flow prediction.Aiming at the problem that the prediction accuracy of traditional Extreme Learning Machine(ELM)in short-term traffic flow prediction is affected by the randomness of input weight matrix and hidden layer bias,a hybrid learning model of PSO-ELM based on ELM is studied.With the help of Particle Swarm Optimization(PSO)in the iterative optimization ability,the global optimal solution of ELM’s input weight matrix and hidden layer bias is found.At the same time,the PSO-ELM hybrid learning model and the traditional ELM model are compared and analyzed at different data sets to reflect the optimization ability of PSO algorithm.Furthermore,the PSO-ELM hybrid learning model is compared with ARIMA,SVR and other classical models on the highway traffic data sets of TDRL and Pe MS to verify the prediction performance.From the perspective of depth temporal characteristic modeling,the short-term traffic flow prediction method based on Temporal Convolutional Network(TCN)model is studied.Through the causal convolutions,dilated convolutions and residual connections in TCN,the depth hierarchy is stacked to realize the deep mining of time series information,the parameter settings and construction of TCN deep learning network model are completed.The experimental results on different data sets show that the prediction accuracy of TCN model is higher than that of PSO-ELM model,ARIMA,SAEs and other commonly used prediction models,which reflects the strong representation ability of TCN model for deep temporal characteristics of short-term traffic flow.Considering the spatial and temporal characteristic of multiple links,the GCN-GRU combined forecasting model based on two-level screening mechanism is studied.Firstly,a two-level screening mechanism composed of autocorrelation function,cross-correlation function and KNN algorithm is constructed to Evaluate the correlation between the target road section and optimize the combination of road sections.Simultaneously,a GCN-GRU combination forecasting model is proposed.The Graph Convolution Network(GCN)is used to capture its spatial characteristics by taking advantage of link topology informations.By adjusting the number of hidden neurons,the long-term memory ability of Gated Recurrent Unit(GRU)is enhanced to extract the temporal characteristics.The prediction result is output through full connection layer.They are verified by the measured short-term traffic flow data of expressway.The results show that using the two-level screening mechanism to effectively screen the road sections and deep learning combination model,the prediction performance will be significantly improved,which is better than the commonly used models such as Stacked Autoencoders network(SAEs)and Temporal Convolutional Network(TCN).