Traffic flow prediction method based on graph neural network

Traffic flow prediction methods, as the key foundation of intelligent transportation systems, are lacking in the existing methods in accurately capturing the spatio-temporal correlation of traffic data. In order to fully explore the complex spatio-temporal correlation of road network and improve the prediction performance, the GAT-TCN model is proposed. In the model, on the one hand, Graph Attention Network (GAT) is used to capture the spatial dependence between nodes in the traffic network, and GAT can adaptively assign weights to the neighboring nodes through the attention mechanism, so that the spatial features under the complex traffic topology can be mined effectively. A discriminator D is added in the training process to optimize the generated time series predictions by means of Generative Adversarial Network (GAN), where GAT in the generator generates predictions, discriminator D discriminates the difference between the predictions and the real data, and ultimately the generator is optimized by fighting against the loss. On the other hand, Temporal Convolutional Network (TCN) is used to process the time series data, by virtue of causal convolution and dilation convolution mechanism, it can efficiently extract the long and short-term temporal features while maintaining the sequence order.The data of each time step is passed through a TCN layer to extract the temporal features, and then these features are fed into the GAT network for graph structure modeling. The fusion of the two enables GAT-TCN to accurately model the spatio-temporal characteristics of the traffic flow data simultaneously.The GAT-TCN model reduces the Mean Absolute Error (MAE), the Root Mean Square Error (RMSE), and the Mean Absolute Percentage Error (MAPE) by 0.11% and 1.45% compared to the existing state-of-the-art models that perform best in PEMS04 and PEMS08, respectively, 2.48% and 6.48% and 4.96% and 5.73%, respectively.