Short-Term Traffic Congestion Prediction Using Public Transportation Data

Abstract

Urban traffic congestion imposes significant costs on commuters, the local economy, and the environment, with its prediction remaining a challenging problem due to the complex spatiotemporal dynamics of traffic flow. In this paper, we present a congestion prediction approach for short-term forecasting that leverages location data from public bus data in the city of Târgu Mures, Romania. The proposed pipeline integrates Automatic Vehicle Location (AVL) data collection, map matching, Travel Time Index (TTI) estimation, and machine learning-based forecasting to deliver 15-minute-ahead congestion predictions. Ten prediction architectures of increasing complexity were evaluated, ranging from naive baselines to recurrent and graph-based deep learning models. The results demonstrate that temporal traffic history is essential for accurate forecasting, and that Graph WaveNet achieves the lowest Mean Absolute Error (MAE 0.546), outperforming models that rely on the fixed road graph structure. The Spatial-Temporal Identity (STID) model achieves superior results on RMSE, MAPE, and R2, representing an efficient alternative. The findings confirm that publicly available AVL data can serve as a reliable and cost effective source for urban traffic monitoring in cities lacking dedicated sensor infrastructure.