Denoising diffusion spatial-temporal residual multi-graph convolutional network for traffic flow prediction

Abstract

Real-world traffic flow prediction remains a significant challenge in intelligent transportation systems. Current methods rely heavily on preprocessed noisy data for model training, yet models trained on such smoothed data often fall short in real-world prediction accuracy. This limitation stems from the overly smoothing effect of interpolation and other preprocessing techniques, which lead to the loss of crucial historical features in the noisy data. To address this challenge, a novel Denoising diffusion Spatial-Temporal residual Multi-Graph convolutional network (DiffSTMG) is proposed for traffic flow prediction. Specifically, DiffSTMG consists of Information Encoding (IE) module, Gated Causal Convolution (GCC) module, Multi-Graph Convolution (MGC) module, Fully Connected (FC) layer and residual connection. The core of the IE module consists of the Denoising Diffusion (DeDiff) module, which is designed to recover the latent features of the noisy data based on diffusion convolution, and the GCC module, which is designed to enhance the temporal features of the noiseless data based on the gating mechanism and the causal convolution. The output of the IE module, after the further enhancement of the temporal features by the GCC module, is passed through the MGC module to obtain the dynamic spatial features. The MGC module is based on the graph convolution network of static and dynamic graphs to further enhance the dynamic spatial features under the influence of historical patterns, locations, and holidays. Extensive experiments on six real-world datasets validate that the DiffSTMG prediction accuracy outperforms state-of-the-art models.