GENERATIVE TRAFFIC FORECASTING: PRESERVING SHOCKWAVE TOPOLOGY WITH DIFFUSION MODELS

Md. Iqramul Hoque, Md Ibrahim Khalil, Tasfia Noor Chowdhury, Tanjim Binta Hasan Jerin

Keywords: Traffic Forecasting, Diffusion Models, Spatiotemporal Modeling, Spectral Bias, Shockwave Topology, Ensemble Medoid Inference

TL;DR: We introduce a diffusion-based generative framework that preserves the physical topology of traffic shockwaves, overcoming the spectral bias inherent in traditional regression forecasting.

Abstract: Standard spatiotemporal graph neural networks for traffic forecasting rely on point-wise regression, minimizing Mean Squared Error. This induces "spectral bias," causing models to act as low-pass filters that smooth out high-frequency, safety-critical shockwaves. We propose a physics-aware generative framework reformulating forecasting as a conditional denoising process. Using a History-Aware Conditional UNet1D and Denoising Diffusion Probabilistic Models, our approach preserves the spatiotemporal topology of traffic flow on the PeMSD7 dataset. To mitigate stochastic variance without reintroducing spectral bias, an Ensemble Medoid inference strategy (N=10) extracts a structurally coherent consensus trajectory. Escaping pixel-perfect MSE minimization incurs an RMSE penalty (9.76 mph) due to the spatiotemporal "double penalty" effect. Nevertheless, our framework achieves a Medoid MAE of 5.32 mph and successfully recovers the sharp kinematic phase transitions and backward propagation of phantom jams that deterministic baselines obliterate, prioritizing physical consistency over pure error minimization.

Submission Number: 135