Research And Application Of Intelligent Traffic State Estimation Based On Attention Mechanism

With the rapid urbanization and the growing increase of the retention of vehicles,the road pressure becomes much higher.The accurate prediction of road traffic states can not only help drivers make better path planning,which saves the travel time,but also assist the traffic management departments in partitioning the traffic network,optimizing the signal timing and guiding the traffic traveling,so as to make full use of road resources and alleviate traffic congestion.In traffic prediction researches,deep learning-based methods attach much more attention in recent years,such as convolutional neural network and recurrent neural network,while most of them do not consider the spatio-temporal characteristic of traffic data.In this paper,traffic data are first divided into several clusters using temporal clustering to the target road to distinguish the traffic environment.Traffic data in each cluster have a similar distribution,which can help improve the prediction accuracy.The hierarchical attention-based model which includes encoder and decoder is applied to capture the spatial features and temporal features respectively.Based on the predicted traffic flow and traffic speed,a fuzzy estimation model is proposed,in which the entropy evaluation is used to set weight coefficients and the fuzzy function is used to calculate membership degrees.Finally,using the fuzzy composition to calculate the traffic state level.The main work of this paper includes:1.Based on the hierarchical clustering,temporal clustering is used to partition the origin data into several clusters.Similarity threshold is set in advance,and the hierarchical clustering is used to merge the most similar clusters,until there’s only one cluster or the maximum similarity among current clusters is lower than the threshold.Finally,Data in the same cluster have similar distribution,and each cluster corresponds to a trained model.2.Using hierarchical attention model to determine the importance of space points and time steps.There are two parts in the model,the spatial attention is used in the encoder to capture the spatial features and determine the importance of each space point with a Bi LSTM network,and the temporal attention is applied in the decoder to capture temporal relations and decide the importance of each time step with another Bi LSTM network.3.Using entropy evaluation to set weight coefficients,and designing fuzzy function to calculate membership degrees.The road’s future state is finally calculated with fuzzy composition.4.Design the traffic state prediction system,which integrates the past traffic data,makes data preprocessing,and visualizes the predicted results.From the practical traffic problems of Hangzhou,this paper completes the research objects of theory analysis,model construction,experimental simulation and system design.Also,some further studies are present to improve the prediction accuracy and integration with traffic environments,so that the reliable traffic states levels can be provided to traffic managements.