Research On Attention-based Intelligent Traffic Flow Prediction Model

With the sustainable development of China’s social economy,the number of motor vehicles has increased year by year,resulting in a series of urban traffic problems such as traffic jams,frequent traffic accidents,etc.With the continuous progress of emerging technologies,intelligent transportation has developed rapidly and gradually entered people’s daily life.Intelligent transportation system(ITS)is a modern means of managing traffic through intelligent information,and traffic flow prediction is the basis and key to the realization of intelligent transportation.It uses massive traffic big data to achieve road traffic flow analysis and prediction of complex urban traffic network,and monitors and collects traffic flow changes in real time through camera networking and computer technology,Through the combination of historical traffic flow data collection and deep learning algorithm,we can predict and evaluate the trend of future traffic flow changes.The prediction of intelligent traffic flow can greatly improve the efficiency of road use,reduce traffic congestion,reduce automobile energy consumption,and improve the traffic capacity of the existing urban road network,which is of great research significance.This topic proposes a short-term traffic flow prediction model for large-scale urban road network to predict the change trend of urban traffic flow in a certain period of time in the future,accurately and effectively select reliable travel modes for travelers,and provide theoretical support for managers to actively implement traffic control.In this paper,three traffic flow prediction models are proposed to solve the problems of the lack of high-quality traffic flow data in the field of traffic flow prediction and the difficulty of capturing the complex spatio-temporal correlation of traffic flow data in the traffic flow prediction network.It mainly includes the prediction model based on data decomposition noise reduction and self attention mechanism,the prediction model based on capsule network and attention module for channel dimension improvement,and the transformation of traffic network into the form of time-space graph to build the time-space feature neural network prediction model based on attention.The main contents of this paper are as follows:1.For traffic flow data are inevitably affected by various external conditions in the real collection process,so this paper proposes an improved integrated empirical modal decomposition algorithm based on the combination of noise reduction decomposition and anomaly processing to remove the noise information in traffic flow sequences;meanwhile,it combines the gated recurrent neural network and the improved self-attentive module to perform the spatio-temporal characteristics and global variation characteristics of traffic data,respectively,from modeling.In the experiment,the combination of anomaly processing algorithm and noise reduction algorithm is used to optimize the noise reduction and training effect of the model and achieve the accurate prediction of traffic flow.2.To address the limitations of traditional convolutional neural networks in capturing the spatial relationships of traffic states in traffic networks,a capsule network is proposed as an improved backbone network to predict traffic flows,which treats traffic networks as graph structures to learn and model spatio-temporal relationships.This chapter proposes to use an improved channel attention and linear bottleneck layer structure in the front-segment feature extraction part of the capsule network to enhance the channel domain attention and information extraction of important features,as well as to enhance the expressiveness of the network.The high-dimensional features extracted in the front segment are transformed into capsule form in the capsule layer,and the information transfer is realized by a routing algorithm to avoid feature loss,and finally the fully connected layer outputs the prediction information.The experimental results show that the improved model is enhanced in traffic flow prediction.3.In view of the limitations of graph convolution network in the application of traffic prediction and the limitation of long-term time dependent modeling capability of traffic prediction based on recurrent neural network,the concept of graph attention network and dynamic graph convolution block are introduced respectively when capturing spatial correlation,and the time trend perception module is introduced when capturing time correlation.The model is based on the framework of basic encoder and decoder,The spatiotemporal sequence prediction of graph structure data is realized by stacking spatiotemporal feature extraction modules in the coding layer and decoding layer.The experimental results show that the traffic flow prediction network stacked by temporal and spatial feature extraction modules has improved in feature extraction and prediction performance compared with other excellent models.