Dynamic Routing Spatial-Temporal Transformer Model for Traffic Flow Prediction

Predicting traffic flow is vital component of Intelligent Transportation Systems (ITS), aimed at enhancing urban traffic management and optimization efforts. However, accurate prediction remains a significant challenge due to the vast array of influencing factors. Many existing approaches often overlook the comprehensive impact of diverse factors on prediction accuracy by considering only a subset of data features. To address this critical issue, a Dynamic Routing Spatial-Temporal Transformer (DRSTT) model specifically tailored for traffic flow prediction is proposed. The DRSTT model captures data features through the design of specialized modules that effectively handle short-term variations, long-term trends, static spatial information of road networks, and dynamically evolving spatial patterns. The proposed approach combines dynamic routing techniques with the transformer model to intelligently adjust information transmission paths within the network according to the distinct features of real-time input data. Utilizing its robust self-attention mechanism, the transformer model effectively captures and analyzes the spatial-temporal dependencies present in data, thereby improving the accuracy of traffic flow predictions. Furthermore, distinct position embeddings are incorporated for different modules to further enhance the model’s capability to recognize and utilize various feature types. The DRSTT model has been thoroughly tested on three traffic flow datasets, consistently surpassing leading techniques and demonstrating its reliability and efficiency in traffic flow prediction.