Physics-Informed Neural Networks for Non-Recurrent Traffic Congestion Detection: A Case Study on the Seoul Ring Expressway

Non-recurrent congestion (NRC), caused by unforeseen events such as crashes, lane closures, and adverse weather, accounts for approximately half of all delays on urban freeways, yet it remains difficult to distinguish from routine congestion at recurrent bottlenecks. This study proposes an NRC detection framework based on a Physics-Informed Neural Network (PINN) that embeds the Lighthill–Whitham–Richards (LWR) conservation law into the learning process to construct a physically consistent baseline of normal traffic states. The traffic flow physics is represented by a two-regime fundamental diagram combining the Greenshields model for free-flow conditions and the Underwood model for congested conditions, and the network is trained by minimizing a composite loss that adaptively balances the data fitting error against the LWR residual. NRC is then detected when the observed density exceeds the PINN-estimated baseline density beyond a tolerance threshold of 150%. The framework was evaluated on a 12 km segment of the Seoul Ring Expressway in Korea using six months of 15 min data collected from seventeen sensor stations. The results show that the proposed model reliably isolates NRC events from recurrent peak-period congestion. From the perspective of symmetry, the framework interprets recurrent traffic as a temporally symmetric background state governed by a conservation law, and non-recurrent congestion as a local breaking of this symmetry, which the physics-constrained residual is designed to expose. The key contribution of this study is a theoretically grounded, label-free anomaly detection approach that couples machine learning with traffic flow theory, offering traffic management centers an automated and interpretable tool for incident detection and response.