Dynamic Correlation-Based Graph Neural Network With Edge-Level Attention for Traffic Congestion Prediction

Abstract

The prediction of traffic congestion is an issue of great significance for the development of intelligent traffic systems for minimizing the delays encountered during travel. In this study, a dynamic correlation-based graph learning framework for efficient traffic congestion prediction is proposed by using a Spatio-Temporal Graph Neural Network (STGNN) and an Edge-Enhanced Graph Attention Network (EdgeGAT). The proposed framework efficiently predicts traffic congestion by using a dynamic graph learning approach that utilizes a sliding window correlation analysis for efficient spatio-temporal dependencies between traffic sensors. In addition, a novel edge representation scheme is proposed based on congestion co-occurrence estimation and free-flow speed estimation for efficient congestion label generation and edge feature learning. The proposed STGNN efficiently learned the spatio-temporal dependencies between the traffic patterns, while the EdgeGAT utilized edge-level attention mechanisms for efficient spatial aggregation for learning the traffic states. In addition to efficient congestion prediction, a congestion-aware rerouting scheme is proposed based on the predicted congestion states for efficient rerouting based on alternative routes with lower traffic density. The proposed framework is tested for efficient traffic congestion prediction using the METR-LA traffic dataset compared to other spatio-temporal graph learning approaches.