Research On Prediction Method Of Urban Road Short-term Traffic Flow

Due to the rapid improvement of economic level and the continuous acceleration of the process of urbanization,the number of private cars in China continues to rise,and traffic congestion has become a common traffic situation in large and medium cities.With the continuous development of science and technology,the research and application of Intelligent Transportation System(ITS)has become important ways and means to solve urban traffic problems,alleviate traffic congestion,and ensure traffic safety.As an important part of Intelligent Transportation System(ITS)research,urban road short-term traffic flow prediction has been widely concerned and researched.Data-driven urban short-term traffic flow prediction methods and models have become an important research direction.This thesis aims to build a short-term traffic flow prediction model for urban roads based on deep learning and machine learning to provide support for real-time refined traffic management and control.Firstly,this thesis classifies traffic flow prediction methods according to different classification standards,and clarifies the definition of traffic flow prediction problems.Aiming at the anomaly and lack of traffic flow data,a three-sigma rule of thumb traffic flow anomaly detection method is proposed,and three methods for complementing traffic flow data are summarized:statistics-based methods,interpolation-based methods,and algorithm-prediction-based methods,and Soft-Impute method based on soft-threshold Singular Value Decomposition(SVD)is used for traffic flow data completion.The temporal correlation,spatial correlation and uncertainty of traffic flow are analyzed with actual traffic flow dataset,and it is clear that the essence of traffic flow prediction is a complex dynamic temporal-spatial sequence prediction problem.Secondly,this thesis summarizes the temporal and spatial feature extraction methods based on deep learning,analyzes the effectiveness of the attention mechanism in traffic flow prediction,and proposes an Attention-based Hybrid Spatio-Temporal Graph Convolutional Network AHSTGCN)traffic flow prediction model,which uses spectral Graph Convolution Network(GCN)in spatial feature extraction,Recurrent Neural Network(RNN)and soft attention mechanism in temporal feature extraction,and draws on the idea of ensemble learning to set up two parallel modules with the same hyperparameters to improve the prediction accuracy and generalization ability of the model.Then a comparative experiment with other methods is carried out on the actual traffic flow dataset,and the traffic flow prediction effect of the AHSTGCN model is analyzed.Finally,this thesis summarizes the online and offline traffic flow prediction methods based on machine learning,analyzes the effectiveness of online prediction methods in solving the concept drift of traffic flow data,a ensemble model of machine learning short-term traffic flow prediction is proposed,which is mainly composed of three modules:offline prediction,online prediction and combined strategy.The online prediction module uses the online AHSTGCN model,Online Passive Aggressive Algorithms(PA)and Hoeffding Tree Regression(HTR),and the offline prediction module uses the offline AHSTGCN model and eXtreme Gradient Boosting Trees(XGBoost).And with the actual traffic flow dataset,the prediction effect of the three combination strategies of more than half simple average,Softmax weighted average and Online Passive Aggressive Algorithms is analyzed.Based on the above research,the real-time prediction model structure of traffic flow in the production environment is proposed.