Traffic Speed Prediction Using MAB-STGNN: Graph Neural Network Built-In Model for Spatial–Temporal Graph Neural Network

Traffic planning, management, and control greatly benefit from accurate and real-time traffic prediction, a critical component of Intelligent Transportation Systems. Despite advancements, traffic speed prediction remains a challenging scientific problem due to the intricacies of topological structures and the multitude of influencing factors that contribute to dynamic changes over time. In order to predict traffic speed more accurately, we propose an advanced model using multi-head attention for spatial–temporal graph neural networks (MAB-STGNN). The proposed MAB-STGNN model is designed as a two-phase model, where we propose a diffusion convolutional layer as the first phase. The diffusion convolutional layer has spontaneous interpretations of the spatial dependencies and can be computed efficiently. The second phase is multi-head attention with a gated recurrent unit (GRU) mechanism, enabling the model to capture diverse features in the spatiotemporal domain. To validate the effectiveness of the proposed approach, we conducted experiments using real-world traffic data from the METR-LA dataset, focusing on the Los Angeles area. The experimental results demonstrate a significant improvement over the baseline methods, showcasing the efficiency of the proposed MAB-STGNN model in improving the accuracy of traffic speed prediction. MAB-STGNN has an enhanced performance in experimental indicators, such as Root Mean Square Error (RMSE), Accuracy, Mean Absolute Percentage Error (MAPE), and Mean Absolute Error (MAE), and shows an ability of long-term traffic speed prediction results.