Traffic Congestion Prediction Algorithms in Urban Environments: A Survey

Traffic congestion poses a significant challenge in urban environments. The use of digital techniques has emerged as a pivotal trend, as it offers substantial safety to and mitigates stress and frustration for road users. The purpose of this survey was to explore the current approaches and digital techniques for managing traffic congestion. We address this through a systematic literature review (SLR) approach by adopting PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. We began by exploring the key techniques of topological data analysis (TDA), machine learning (ML) and deep learning (DL) for modeling urban traffic prediction. We evaluated the robustness of the topological data analysis technique (Persistent Homology (PH)) against deep learning frameworks (Graph Convolutional Neural Networks (GCNNs)). We found that each framework has its own strengths and weaknesses, and neither of the frameworks independently provides a complete solution. PH may offer richer structural insights and robustness to noise but may struggle with direct predictive implementation, while deep learning models do better at extracting dynamic predictive patterns but are assumed to lack interpretability and generalizability. Therefore, the integration of multiple techniques, either PH with stacking ensemble methods or deep learning with stacking ensemble methods, can improve prediction and generalization of the model while at the same time reducing over-reliance on local graph assumptions. Future research should focus not only on performance metrics or methods but also on explainability, transferability, adaptability across heterogeneous road environments and computational cost.