GNN-Based Urban Congestion Prediction Considering Congestion Prior Knowledge
TL;DR
This model provides a novel solution for traffic congestion prediction in ITS that balances physical interpretability with predictive performance, ultimately contributing to more reliable and practical intelligent transportation management systems capable of addressing real-world congestion challenges with improved accuracy and robustness.
Abstract
Traffic congestion prediction offers a proactive perspective for alleviating urban traffic congestion and represents a critical task within intelligent transportation systems (ITS). To bridge the gap between traffic congestion prediction models and domain-specific prior knowledge, this paper introduces a congestion-prior mixed graph convolutional recurrent network (CMGCRN) that explicitly incorporates traffic congestion prior knowledge, thereby effectively boosting the accuracy of traffic congestion prediction. First, we design an ST-Apriori algorithm incorporating graph constraints and temporal constraints to mine spatiotemporal correlation patterns between the traffic congestion phenomenon and urban road network structures. Then we propose a novel graph construction method: the local complete graph, which explicitly encodes the physical propagation characteristics of recurrent traffic congestion in local areas. Subsequently, CMGCRN is developed by organically integrating data-driven dynamic graph, global predefined graph, and local complete graphs to establish a multidimensional fused urban traffic congestion prediction model. Experimental results demonstrate that, under peak-hour conditions, CMGCRN achieves respective improvements of 1.7, 0.4, and 1.7% in F 1 -scores for congested status prediction across three different time horizons compared to the suboptimal baseline model, effectively enhancing congestion prediction accuracy. Further analysis reveals that the model maintains performance stability in long-term predicting scenarios while significantly reducing critical misclassification errors, particularly in congested state identification. By deeply integrating knowledge guidance with data-driven approach, our model provides a novel solution for traffic congestion prediction in ITS that balances physical interpretability with predictive performance, ultimately contributing to more reliable and practical intelligent transportation management systems capable of addressing real-world congestion challenges with improved accuracy and robustness.