Traffic prediction based on spatio-temporal feature embedding fusion and gate operation optimization

Abstract

In real-world traffic prediction problems, there are often complex spatio-temporal features and patterns. To enhance the accuracy and performance of traffic prediction and address these complexities, it is essential to employ effective models and methods to capture spatio-temporal features and patterns of change. For this purpose, we propose a network model that integrates spatio-temporal feature embeddings with gate operation optimization(TSGO). In our model, we design a novel module: the spatio-temporal feature embedding fusion module, which combines input data to strengthen the model’s ability to extract spatio-temporal correlation features, particularly in enhancing temporal features. To further bolster the capture of spatial features, we design an adaptive graph structure learning method based on a node repository, dynamically capturing non-Euclidean spatial correlations within the traffic network. Additionally, to better capture long-term dependence and short-term variations in sequential data, we adopt a new strategy in the Gated Recurrent Unit (GRU): treating the even and odd positions in the input sequence as two separate input streams to generate corresponding update gates and reset gates. This approach enables the model to utilize data more evenly, achieving complementarity between the two sets of features and allowing it to adapt to information at different time scales within the sequential data. In short-term, medium-term, and long-term predictions across three real-world traffic datasets, the TSGO model achieved average MAE reductions of 8.76 %, 10.12 %, and 11.86 %, respectively, compared to the baseline. This demonstrates its capability to generalize across different time scales and significantly improve prediction performance.