T-GANRE: Advanced Deep Learning Models for Synthetic Data Generation for Nonrecurrent Events
Collin Meese, Danielle Lee, Kurt Hammen, Mark Nejad · 2025
Nonrecurrent events (NREs), such as vehicular accidents and work zones, are a substantial challenge in accurate and reliable near-real-time traffic prediction and management due to their sporadic nature and unique and complex influences on the traffic network. Existing state-of-the-art deep learning (DL) methods often inadequately or fail to capture the nuances of these events due to the sparsity of NRE data, which is imbalanced with the abundant recurrent data, often resulting in the model considering these events as noise. To address these limitations, we propose a novel approach for leveraging a Generative Adversarial network (GAN) with a transformerbased architecture. Our method generates high-quality synthetic traffic time series data for NRE scenarios, providing models with sufficient robust representation of post-incident traffic dynamics. We amalgamated open-source traffic data to generate labeled NRE metadata that our model uses to generate synthetic sequences conditioned on pre-incident traffic states. Comprehensive experimental evaluation demonstrates the synthetic data's realism, aligning firmly with real-world metrics, including temporal patterns and distribution similarity. This model and subsequent synthetic dataset bridge the gap in NRE representation and enhance understanding of NRE impacts, providing a foundation for advancing proactive traffic management strategies.