Zhang, Dingkai (Tongji University), Wang, Pengfei (East China Normal University), Ding, Lu (Tongji University)

Generative Adversarial Network-Based Adaptive Spatio-Temporal Traffic Flow Prediction with Large Language Models

Scheduled for presentation during the Invited Session "S23a-Trustworthy AI for Traffic Sensing and Control" (TH-LM-T23), Thursday, November 20, 2025, 11:10−11:30, Coolangata 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Accurate traffic flow prediction is a critical component of intelligent transportation systems. While previous studies have made significant breakthroughs in modeling spatio-temporal correlations, existing methods still suffer from two significant limitations: (i) underdetermined traffic flows caused by sparse sensor deployment, where incomplete path observations make it difficult to capture non-local spatial dependencies. (ii) non-equilibrium traffic flows arising from congestion propagation and varying driver reaction times, which induce asynchronous temporal correlations and nonlinear dependencies. To address these challenges, we propose a novel Generative Adversarial Network and Large Language Model-based Adaptive Spatio-Temporal framework (GLAST). Specifically, It employs an adaptive edge weight augmentation mechanism based on GAN with residuals to capture deviations in periodic patterns and construct dynamic adjacency matrices, which enables the capture of non-local spatial dependencies under sparse sensing conditions. The weighted attention mechanism is applied to aggregate lagged adjacency matrices to model asynchronous temporal correlations, effectively capturing temporal delays and nonlinear dependencies. Furthermore, adaptive temporal descriptors and congestion metrics are encoded as prompts for the LLM to enhance the modeling of multi-scale spatio-temporal relationships. Experimental results on benchmark datasets demonstrate that GLAST consistently outperforms various state-of-the-art baselines. Given the widespread presence of sparse sensing and congestion propagation in real-world highways, the proposed framework also holds the potential to inspire other urban traffic applications.