| With the rapid development of China’s economy,vehicle ownership is increasing quickly,which has already caused a series of problems such as traffic congestion,and brought great inconvenience and distress to travelers and managers.Intelligent transportation systems provide an effective means to alleviate traffic congestion.Traffic flow prediction is the foundation and key to realize intelligent transportation systems.Using big data to achieve traffic prediction for road networks is of great value in theory and practice.This thesis proposes several short-term traffic flow prediction models for large-scale urban road networks,providing travelers with accurate travel information and managers with recommendations of active traffic management.The main research contents and results are summarized as follows:(1)Theoretical and methodological analysis for short-term traffic prediction.We analyze basic parameters of traffic flow and their characteristics,and propose some capabilities prediction models should have.The theoretical basis for the establishment of traffic flow prediction models is provided through deep theoretical analyses.(2)Prediction of traffic flow on expressways.We choose some important traffic flow parameters and analyze their spatial-temporal correlations.Based on convolutional neural networks,a novel traffic flow prediction model is proposed for expressways.The model can capture spatial-temporal correlations of various features and simultaneously predict multiple traffic flow parameters.We then use real-world data to train and test the model.(3)Predict urban network-wide traffic flow considering spatial-temporal correlations.A deep tensor is constructed,and a novel model called spatial-temporal deep tensor neural network(ST-DTNN)for road network-wide prediction is proposed.The potentially negative effects caused by the manually stacking order of time series are eliminated.At the same time,parameter updating algorithms are deduced and designed for model training and learning.A case study was performed on the model using real-world large-scale urban road network data.The results show that the ST-DTNN model greatly improves the prediction performance compared to traditional time series and machine learning benchmarks.(4)Road network-wide traffic prediction considering network topology.A traffic flow prediction model based on graph convolutional network(GCN)is constructed.The topological structure of the road network is considered as well as the spatial-temporal correlation of traffic flow,which improves the prediction accuracy and interpretability of the model.Compared with the ST-DTNN model,it is found that the GCN model has a certain improvement in prediction accuracy,while the training time is shortened by nearly 35%,and the model becomes more efficient.The GCN model further improves the traffic flow prediction framework on the urban road network level based on ST-DTNN.Its high-efficiency and high-precision characteristics enable the applications to real-world traffic management and route planning. |
