| With the development of intelligent transportation system and data infrastructure,along with the huge increase in vehicle traffic,huge amounts of traffic flow data have also been generated.The prediction of short-term traffic flow is the basis and important application component in the frame of ITS.In the research of intelligent transportation systems and traffic big data,short-term traffic flow prediction has always occupied a larger proportion.For the short-term traffic flow prediction model,the current starting point is how to learn the characteristics of the traffic flow more comprehensively,and avoid the performance loss to improve the prediction performance of the model.Aiming at the existing performance loss of existing deep neural network traffic flow prediction models which caused by overfitting.A regularized LSTM model that fused recurrent Dropout and max-norm weight constraint was propose in this dissertation.Before the prediction model is established,the Canopy-K-means model is used to improve the general K-means traffic clustering algorithm,and the traffic areas with similar temporal characteristics are concentrated,which is more conducive to model identification.And the correlation coefficient formula is introduced to calculate the correlation of neighboring stations,the highly relevant station traffic is introduced as an additional feature.For the regularized LSTM model,the recurrent Dropout apply to the recurrent connection,the maximum normal constraint will limit the input weights.The whole model uses Adam optimization algorithm to optimized.Through experiments on three real traffic datasets from different countries and region,it is proved that the regularized LSTM model improves the prediction performance compared with the basic LSTM model and other commonly used methods.At the same time,compared with other researchers on the same PeMS dataset,the results proved that our model performs better on performance indicators. |
PDF Full Text Request