Tail-aware spatiotemporal diffusion for probabilistic forecasting of extreme traffic congestion

Extreme congestion occupies only a small portion of traffic records but carries disproportionate operational consequences for traffic-management systems. Conventional point and probabilistic forecasters are optimised primarily for average accuracy and therefore often underfit rare, severe speed reductions. We present Tail-Aware Spatiotemporal Diffusion (TAD-ST), a probabilistic forecasting framework that couples a temporal convolutional and graph convolutional encoder with a conditional diffusion decoder. A congestion-aware denoising objective assigns greater training influence to future windows containing severe congestion without resampling the training data. The framework forecasts the probability and evolution of speed-derived extreme congestion from historical traffic states and network structure. Evaluation on METR-LA and PEMS-BAY uses training-only congestion thresholds, overall and event-conditional probabilistic metrics, multiple random seeds, component ablations, calibration, transferability, and inference-efficiency analysis. In the reported comparisons, the largest separation from the architecture-matched uniform-loss baseline occurs on event-conditional CRPS, whereas average point accuracy remains competitive. The evaluation therefore emphasises calibrated tail risk rather than uniform dominance on every metric.