Research On Urban Traffic State Prediction Driven By Multivariate Data

With the rapid development of information technology,intelligent transportation construction has become a new direction for the development of many cities.The upgrading and optimization of Intelligent Transportation System(ITS)is the main content of intelligent transportation construction,and urban traffic state prediction is one of the key technologies to realize traffic control and guidance in ITS.A traffic dataset with complete information and high value density can effectively improve the accuracy of traffic state prediction,assist traffic management department to monitor the overall traffic operation state in real time and make correct decisions.The existing traffic datasets are mainly collected by one-way lane of expressway,and less consideration is given to other factors affecting traffic operation.Due to the lack of traffic auxiliary information,the accuracy of traffic state analysis using such dataset is not high.The spatiotemporal characteristics of traffic data itself are complex,the existing traffic prediction models only analyze the temporal or spatial characteristics of the data,and the spatial-temporal feature extraction of traffic data is not complete,which results in the problems of low precision of prediction model and low efficiency of actual scene utilization.In view of the problems existing in the above traffic state prediction,this paper obtains the real-time road information of many cities in Amap with network crawler technology,including 308 road traffic data of Xi’an,51 traffic state data of Xining City and traffic data of8 main sections in Beijing.In addition to the traffic state data,the obtained dataset also involves the regional weather data for 2 minutes,with 17 main parameters.Compared with the other ones available at present,the datasets obtained in this paper have a long span,large data volume and more complete information.It can provide effective data support for the urban traffic state prediction when it is true.This paper first preprocesses the acquired datasets,including redundancy processing,missing value filling,and semantic transformation.Then,analyze the own characteristics and correlation of the parameters in the dataset,and obtain the characteristics of the data itself and the correlation degree between the parameters,so as to facilitate the subsequent model selection and data dimension reduction.Then the deep study model is used to analyze the multiple parameter combination of different time periods(workday and weekend)to explore the impact of different parameters combination on traffic prediction and the sensitivity of the model to traffic data in different time periods.Finally,this paper proposes a bi-directional long-and short-time memory neural network model based on the attention mechanism(Attention mechanism Bi-directional Long Short-Term Memory,ABi-LSTM),comprehensively considering temporal and spatial characteristics.and use the memory unit in the Bi-LSTM to learn the time characteristics of data,learn the spatial characteristics of Bi-LSTM and realize adaptive attention on features with the attention mechanism.The model,on the basis of considering the time characteristics,considers the influence of vertical multivariate data(upstream and downstream sections)on the traffic state prediction,realizes the allocation of two types of characteristics weights combined with the attention mechanism,and solves the low accuracy of incomplete spatial-temporal feature extraction of traffic data and not applicable to the actual traffic state evaluation.Experimental results show that the Mean Absolute Percentage Error(MAPE)of traffic state prediction based on multivariate data is nearly 30 percentage points lower than that of single data traffic state prediction.At the same time,the prediction results of the model under weekday and weekend are in line with the actual observation law and ideal.Compared with the experimental results without spatial characteristics,the MAPE of the ABi-LSTM model considering spatial-temporal correlation can be reduced by 20 percentage points.Compared with the datum model,the proposed ABi-LSTM model considering spatial correlation has high accuracy in traffic state prediction.