Study On Short-term Traffic Flow Velocity Prediction Model Based On Attention Mechanism

Intelligent transportation system(ITS)plays an important role in solving a series of traffic problems,and short-term traffic flow prediction is the core content of Intelligent Traffic System and the important foundation of traffic information service,control and guidance system.Deep learning algorithm can use multi-layer neural network or deep architecture to capture the inherent characteristics of data,and the improvement of deep learning algorithms performance by introducing attention mechanism has been verified in natural language processing,image recognition and other fields.Due to the complex and random characteristics of traffic flow,accurate identification of traffic flow characteristics is not a simple task.It is very important for the whole traffic system to use the deep learning algorithm to predict the traffic flow more accurately.In this thesis,the deep learning algorithm is used to study the short-term traffic flow velocity prediction of the road network.In view of the shortcomings of the existing research results,this thesis proposes two specific methods from the perspective of prediction accuracy and prediction model efficiency.The research content of this thesis includes:(1)In order to improve the quality of the data of used to predict,the thesis analyzes the statistical characteristics of traffic flow data,carried on the system’s data preprocessing,gives the definition of traffic flow data missing divided,and according to the definition of accidental missing and multiple missing respectively using Naive Bayes and Dynamic Time Warping algorithm to estimate and fill the data.(2)This thesis expounds the reason for choosing the Recurrent Neural Network as the basis for traffic flow prediction,and puts forward the Long Short-Term Memory to improve the model for its gradient disappearance in practical application,and then describes the model framework,algorithm and training process in detail.(3)In order to further improve the performance of the model,the attention mechanism was introduced into the model,and the variant based on the attention vector calculation method adapted to the research scene of this paper was selected.The short-term traffic flow velocity prediction model based on the attention mechanism was constructed,and then the improved model framework,neural network structure and training process are described.(4)The built model is verified by using the actual data of the urban road network.The results show that the hybrid model proposed in this thesis has higher repair rate and lower estimation bias than the model without data missing partitions.Compared with deep learning algorithms such as RNN and CNN,The proposed prediction model in this thesis can improve prediction accuracy and model efficiency.