Research On Traffic Big Data Repair And Prediction Based On Deep Learning

The development of advanced information technology and electronic sensing technology provides massive data support for the deployment of Intelligent Transportation Systems(ITS)and theoretical research in the field of transportation.How to effectively mine potential traffic dynamic regularity from the collected traffic data and serve traffic participants through ITS is a hot topic of research.Due to the large amount of traffic data collected and various modes,it is challenging to establish a reliable traffic flow analysis model based on traditional probability theory,fluid mechanics,and dynamics.Deep learning(DL),as a new and advanced data-driven machine learning method,has aroused great interest in academic research and industrial applications.The use of deep learning to impute missing traffic data,predict traffic congestion evolution trends,and determine congestion nodes are of great significance for tasks such as next-generation traffic control optimization and government ease congestion.This thesis focuses on the problem of imputing missing traffic data and predicting traffic flow based on deep learning.The main research work and innovations are as follows:(1)In view of sensor failure,communication error,storage loss,etc.,the data collected by the sensor will inevitably have data loss problems.An algorithm for imputing missing traffic data based on self-attention and adversarial auto-encoder(SA-AAE)is proposed.The proposed method can effectively capture different correlation weights between input data and generate imputed data close to real data distribution.The effectiveness of the method proposed is verified by comparison with different imputation models.(2)A short-term traffic flow prediction model based on spatio-temporal analysis and Convolutional Neural Network(CNN)is proposed.In the proposed model,the best input data is determined by the Spatio-Temporal Feature Selection Algorithm(STFSA),and the complex dependencies in the input data are learned using CNN to build a predictive model.The effectiveness of the proposed method is verified by simulation comparison with other typical traffic flow prediction models.(3)A short-term traffic flow prediction model based on time series decomposition analysis and Long Short-term Memory Neural Network(LSTM)is proposed.In the proposed model,the periodicity,short-term dependence,and heteroscedasticity volatility of traffic flow sequences are analyzed,and combined with LSTM and Deep Neural Network(DNN)to obtain higher prediction accuracy.Comparative analysis with classical statistical methods and the latest neural network model shows that the proposed algorithm is an effective short-term traffic flow prediction algorithm.