Regime-aware congestion diagnosis and prediction using physics-informed tensor learning

Existing learning-based traffic prediction models achieve high accuracy in short-term speed and flow forecasting, yet they rarely elucidate the underlying causes of traffic congestion, thereby offering limited decision-support for selecting optimal mitigation strategies. Applying the wrong traffic solution wastes resources while allowing congestion queues to grow. This article closes this gap by turning a congestion observation into an actionable diagnosis. We reformulate recurrent and non-recurrent congestion as physical regimes characterized by interpretable Newell signatures, such as the start time of congestion period t0, the time index with maximum queue length t2, the congestion duration P, the discharge rate μ, the queue length Q, and the demand-over-capacity D/C, and decompose the non-recurrent deviation into three distinct categories: demand-side deviation, supply-side deviation, and phase-boundary deviation. A Newell-residual signal combined with a Random Forest (RF) classifier separates the recurrent baseline from non-recurrent observations. The observed data are organized into a segment-by-time-by-feature tensor, and a physics-informed multi-view clustering method (Stepwise Common Principal Components with k-means [SCPCk]) groups physically homogeneous road segments. A group-sparse tensor model reconstructs traffic states and a fluid queue model converts each diagnosis into an interpretable estimate of congestion duration. Numerical experiments on six synthetic datasets and a real PeMS_Bay (part) corridor dataset show that the proposed framework achieves competitive state-reconstruction performance against DCRNN, PLS, and HA baselines while substantially improving structured-deviation identification and yielding congestion-duration estimates with the lowest errors among baselines.