Generative AI for traffic scenarios: a GCN-VAE model

This paper presents a novel generative model based on the Variational AutoEncoder (VAE) architecture, designed to generate realistic road speed scenarios in road networks. The proposed approach incorporates Graph Convolutional Networks (GCN) to effectively capture spatial correlations inherent in the road network topology. Unlike traditional statistical methods which often require prior knowledge of data distributions and substantial computational resources, the architecture presented in this work provides a more efficient solution, capable of generating high-quality synthetic data. Experimental evaluations on a real-world dataset demonstrate the model’s ability to accurately replicate real traffic data distributions. These findings highlight the potential of combining VAEs with graph-based techniques to significantly enhance urban traffic simulation and planning.