Traffic Flow Analysis And Prediction Based On Data Mining Technology

With the increasing number of motor vehicles,traffic congestion has become a major problem that constrains social development.If the traffic flow of each intersection in the road network can be predicted more accurately,the authorities can make more targeted measures.In this way,the road network utilization can be improved comprehensively to alleviate traffic congestion.In recent years,with significant advances in data mining technology,the combination with traffic flow prediction scenarios is promising.In this paper,the traffic flow is analyzed and predicted based on data mining technology.Firstly,based on the clustering technology,a mining method for traffic flow pattern is presented.By calculating the three coefficients of the K-Means and Agglomerative Clustering algorithm,i.e.the silhouette coefficient,Calinski-Harabaz index and Davies-Bouldin index,the amount of the traffic flow pattern is determined.Thus,the traffic flow patterns are divided into two types:single peak pattern and double peak pattern.What's more,it is verified that holidays and working days correspond to different traffic patterns respectively.Secondly,two traffic flow prediction models based on deep learning are proposed corresponding to two different kinds of raw data.On the one hand,for the data provided by only one monitoring station information,the traditional single recurrent neural network cannot deal with the irrelevancy after information splicing.In order to solve this problem,the Combined-Update Gate Recurrent Neural Network(CUGRNN)model is presented to extract the temporal characteristics of traffic information with multiple recurrent neural networks.On the other hand,for the data providing information not only one monitoring station but also upstream and downstream information,the complexity of CUGRNN increases dramatically as the input sequence increases.In addition,the CUGRNN cannot learn the spatial relationship among stations.In order to solve this problem,the Convolutional-UGRNN(Conv-UGRNN)model is presented to simplify the model and to capture the spatial relationship by convolutional neural network.Thirdly,in order to solve the problem of limited traffic data volume,a data augmentation method for traffic information is realized with convolutional auto-encoder.Considering the internal relation among different traffic information,the data volume is reconstructed and extended.Thus,the CUGRNN and Conv-UGRNN can learn more traffic flow information and show higher accuracy.The experimental results show that the division of traffic flow pattern improves the purity of training set and reduces the difficulty of fitting.Besides,the CUGRNN and Conv-UGRNN are more accurate than the algorithms proposed in the literatures thus their network structures are considered more reasonable.Last but not the least,the data augmentation method based on convolutional auto-encoder improves the accuracy of the two models effectively.