Graph Neural Network Based Spatial-Temporal Traffic Flow Prediction Approaches
MetadataShow full metadata
The problem of traffic flow prediction is an important topic in the research of Intelligent Transportation System (ITS). With the acceleration of urbanization, the pressure of traffic load has increased. These situations urgently require scientific management and scheduling. Therefore, the development of intelligent transportation systems is imperative. The maturity of machine learning technology and the continuous development of graph neural networks allow us to better understand the temporal and spatial dynamics features of traffic flow data hidden in complex traffic networks. However, this is very challenging because of the high degree of nonlinearity, complexity, and randomness of traffic flow. These factors make the traffic flow difficult to predict and lead to low prediction accuracy, which is difficult to meet the needs of application scenarios. Traditional traffic flow prediction models and methods lack the ability to extract periodic characteristics of traffic flow data, which makes it impossible to learn more powerful traffic flow feature data reasonably. Moreover, many existing machine learning models do not fully consider the correlation between the traffic flow sequence in the spatial dimension and the temporal dimension, which makes the general applicability of the models insufficient. In addition, most combined deep neural network models ignore the characteristics of the traffic network graph structure and cannot express the high-order correlation between different nodes. In response to the above problems, this thesis proposes four Graph Neural Network-Based Spatial-Temporal Traffic Flow Prediction models to improve the accuracy of traffic flow prediction further. First of all, this thesis adopts reasonable data analysis and dimensionality reduction strategies to improve the reliability of input data and reduce its complexity. These methods improve the model's ability to extract traffic flow features in the data input stage. Secondly, based on the Graph Neural Network (GNN), our models improve the interpretability and accuracy of the models in the temporal dimension through the advantages of Gate Recurrent Unit (GRU) and Temporal Convolutional Network (TCN) in the temporal dimension feature processing. Combined with the powerful extraction capability of the graph convolutional neural network module for spatial dimensional features, the models' general applicability and prediction accuracy are enhanced. On this basis, this thesis uses the self-attention mechanism to enable the models to capture the dynamic dependence of traffic flow data in temporal and spatial dimensions, thereby further improving the prediction accuracy of the models. In this thesis, the models are tested on two real traffic flow data sets. The simulation results confirm that the models can be effectively used for traffic flow prediction, and the prediction accuracy is better than other similar methods. Especially when the prediction step is long, the models have more obvious advantages in prediction accuracy. Due to the advantages of the GRU module in sequence data processing and the ability of the attention mechanism to extract dynamic dependencies between nodes, the MST-AGCRN model has higher prediction accuracy than other models we proposed. At the same time, the MST-AGCTN model has higher complexity and more parameters, and its performance is lower than expected. It needs further research and exploration.