Benchmarking/Limitations of Traffic Prediction with Noisy Field Measurements

Traffic prediction is essential for urban planning, congestion control, and emergency response. However, the presence of noisy and incomplete field measurements poses significant challenges to accurate forecasting. This study assesses the performance and limitations of traffic prediction models under noisy data conditions, introducing a novel incomplete sequence training approach to improve robustness with incomplete data. Our method is assessed on a comprehensive Las Vegas road network, revealing that the graph-based spatio-temporal transformer model consistently outperforms alternative approaches. These results underscore the model's ability to generate reliable predictions, thereby enhancing its practical applicability in real-world scenarios with messy data. This research emphasizes the critical role of robust training techniques in mitigating the adverse effects of noisy field measurements on traffic prediction accuracy.