Spatiotemporal expert GPT network for traffic flow prediction

Abstract

Urban traffic displays pronounced temporal periodicity, spatial dependencies, and heterogeneous behavioral patterns. These characteristics hinder traditional approaches from capturing multi-scale dynamics and complex topologies within a unified prediction framework. To address this challenge, a Spatiotemporal Expert GPT Network (STE-GPT) is introduced as a unified paradigm that formulates traffic flow prediction as a language-style sequence modeling problem. The framework employs a Spatiotemporal Encoder to transform node observations, temporal context, and road-network topology into unified spatiotemporal token representations. This design enables a principled shift from graph-signal learning to semantic-level sequence modeling. Based on this representation, Sandglass Attention conducts topology-aware spatial compression and reconstruction. This process improves the efficiency of global dependency modeling. In addition, the LLM-Guided Expert System integrates a pretrained Transformer with a sparse expert mechanism. A gated routing strategy activates specialized subnetworks based on traffic conditions, enhancing reasoning capacity under heterogeneous patterns. In experiments conducted on four real-world datasets, the STE-GPT model proposed in this paper achieved improvements of 8.66% and 1.03% in MAE and RMSE, respectively, and 4.57% and 0.34% in MAE and RMSE, respectively, compared to the current state-of-the-art baseline method, STLLM, on the PEMS07 and PEMS08 datasets.