TRL-Trans: Adaptive Graph Learning with Temporal Representation for Traffic Flow Prediction

Traffic flow forecasting plays a pivotal role within intelligent transportation frameworks. Although existing methods combine graph neural networks and temporal models, there are still problems, such as static graph structure being challenging to characterize the dynamic associations between traffic nodes, insufficient ability to model long temporal dependencies, and low efficiency of fusion of complex spatio-temporal features, etc. Based on this, we propose a Transformer-based Temporal Representation Learning traffic flow prediction model (TRL-Trans). The proposed model employs Temporal Representation Learning (TRL) to derive contextual insights from heavily masked subsequences. It incorporates a Gated Temporal Convolutional Network (Gated TCN) coupled with an Adaptive Hybrid Graph Convolution Module (AHGCM) to effectively capture dynamic spatio-temporal characteristics. The AHGCM dynamically merges predefined adjacency matrices with implicit spatio-temporal relationships. Additionally, the Transformer component strengthens the model’s capacity to handle extended temporal dependencies. The experiment shows that TRL-Trans outperforms the baseline model.