Traffic Time Series Analysis And Forecasting Based On Complex Networks And Deep Learning

In order to solve the increasingly serious traffic problem and develop sustainable transportation,the concepts of intelligent transportation system and smart transportation have been continuously proposed,and have been explored and practiced in recent years.With the development of software and hardware technologies such as data acquisition and artificial intelligence,how to use massive traffic data to mine the inherent laws of complex transportation systems has become a key issue.In recent years,the collection,analysis,mining and prediction of traffic data have gradually become the research focus in the field of traffic,and played a key role in the refinement of traffic organization and guidance.Based on the above background,this paper takes traffic time series data as the research object,and uses machine learning,complex network theory,deep learning and other methods to repair,mine and predict the data respectively.The main work is as follows:Firstly,the application of complex network in traffic field,analysis of time series data and short-term forecast of traffic flow at home and abroad are combed and expounded.On this basis,the research direction of this paper is determined and the research route of this paper is designed.Secondly,an improved data preprocessing method is proposed.Traffic system is a complex,open and uncertain system,and the collected traffic data are often affected by the detection equipment,collection environment,traffic events and other factors,resulting in the absence and noise.Based on the consideration of data characteristics and traffic conditions,this paper proposes a univariate data filling method based on clustering algorithm and a multivariate data filling method based on residual prediction.Thirdly,the complex network of traffic time series data is constructed and analyzed.Phase space reconstruction(PSR)and adaptive network density method are introduced to construct the traffic time series data into a complex network.This paper verifies the scale-free characteristics of the network,analyzes the correspondence between the structure of the network and the traffic state,analyzes the distribution relationship between the clustering coefficient and the betweenness of the network nodes,quantitatively gives the identification method of the important nodes in the network,and on this basis,the key nodes are mapped to the key time window of the corresponding traffic time series,which can provide a reference for the implementation of refined traffic control measures.Finally,a short-term traffic forecast model is constructed.This paper expounds the algorithm structure of long-short term memory neural network(LSTM)in detail,and discusses the superiority of this algorithm structure in processing time series data,which lays a theoretical foundation for modeling and forecasting.Combining phase space reconstruction with long-short term memory neural network,a PSR-LSTM short-term traffic prediction model is constructed,and the network structure and prediction process of the model are explained in detail.Select typical machine learning algorithms such as support vector regression,extreme gradient boosting tree and multi-layer perceptron to build two groups of control experiments,and make speed prediction and travel time prediction respectively.The experimental results show that the PSR-LSTM short-term traffic prediction model proposed in this paper is superior in many precision indexes such as average absolute error,and the prediction accuracy of the model is higher.