Self-supervised short-term traffic flow prediction based on diffusion autoregressive Transformer

Abstract

With rapid urbanization and motorization, traffic congestion restricts urban sustainable development, yet existing short-term traffic flow prediction methods have flaws: Linear Regression fails to adapt to nonlinear features like peak tidal changes; SVM has soaring kernel cost in large-scale data and poor long-term dependency capture; LSTM lags in extreme loads and has high error fluctuation in special events. This study proposes the STD-ARformer, integrating diffusion denoising and autoregressive mechanism. It uses three core designs: dynamic receptive field (adjusts attention window for mutations), traffic flow conservation constraint (ensures physical compliance), and hierarchical denoising (enhances multi-scale robustness).Experiments show STD-ARformer outperforms Linear Regression, SVM, and LSTM in key indicators, alleviates extreme load lag, reduces special event error fluctuation, and lowers medium-flow discreteness. It provides a high-performance solution, supporting traffic management and urban sustainable development.