Adaptive Anomaly-Residual Early Warning System for Proactive Detection of Urban Traffic Congestion
Urban traffic congestion remains a major challenge for metropolitan transportation systems because it increases travel delay, fuel consumption, and environmental pollution. Conventional monitoring systems are mostly reactive and report congestion after traffic breakdown has already occurred. This paper proposes a lightweight Adaptive Anomaly-Residual Early Warning System (AAR-EWS) for proactive congestion detection. The framework integrates a neighbor-aware Temporal Convolutional Network (N-TCN), residual-based anomaly scoring, segment-wise adaptive thresholding, uncertainty-gated alert logic, and propagation-aware neighborhood confirmation. Experiments on the METR-LA benchmark show that the proposed method reduces 15-minute MAE by 6.3% compared with the TCN baseline, provides an average congestion lead time of 13.6 minutes, and reduces false alerts by more than 35% compared with threshold-based warning rules. The revised study further clarifies model configuration, operational assumptions, robustness under noisy and missing data, deployment feasibility, and comparison with graph neural network approaches.