A Spatio-Temporal Dynamic Graph Model Based on Dual-Stream Convolutional Neural Network for Traffic Speed Prediction

Accurate traffic speed prediction is crucial for advancing intelligent transportation systems in urban environments. However, traffic speed is influenced by a multitude of factors, such as traffic volume. Although some studies have attempted to integrate these variables into their models, the underlying mechanisms governing their interactions remain insufficiently elucidated. This lack of clarity impedes the ability to identify latent patterns embedded within traffic flow data. Furthermore, the nonlinearity and high-dimensionality of traffic speed data, coupled with the dynamic nature of traffic networks, often constrain the predictive performance of standalone models, rendering precise traffic speed prediction a formidable challenge. To address these challenges, we leverage a dual-stream convolutional neural network (DS-CNN) to explore the high-dimensional relationships between traffic flow and speed. Additionally, by employing an adaptive graph learner (AG), the model dynamically updates the traffic network graph during training, thereby effectively simulating real-world traffic scenarios. Moreover, by substituting linear transformations within a gated recurrent unit (GRU) with graph convolution (GCN) operations, the model simultaneously captures both temporal and spatial dimensions, thus uncovering intricate spatio-temporal dependencies. We name the proposed hybrid model DSAG-GCGRU. Experimental evaluations conducted on the widely recognized PEMS series datasets demonstrate that our approach outperforms state-of-the-art baseline models in medium- to long-term predictions, achieving average improvements of 1.56% in mean absolute error (MAE) and 7.88% in mean absolute percentage error (MAPE). Therefore, the model proposed in this study is highly suitable for medium- to long-term traffic speed prediction, providing managers with effective decision-making insights.