Probabilistic Spatio-Temporal Traffic Flow Forecasting Based on Denoising Diffusion Framework

Probabilistic spatio-temporal traffic flow forecasting holds significant value in intelligent transportation systems. Recently, research based on denoising diffusion models is booming and has become one of the most promising approaches to deal with this problem. Despite certain achievements, most existing studies use static spatial graphs based on prior knowledge to model spatial dependencies, which makes it difficult to effectively capture the complex spatial dependencies in real traffic networks. In addition, they do not fully utilize the intrinsic features of spatio-temporal traffic flow data. To address these issues, we propose a novel conditional denoising framework called STDiff, which can utilize spatio-temporal information and intrinsic characteristics in traffic flow data better. Specifically, our method effectively captures spatio-temporal dependencies through MLP and spatio-temporal attention mechanisms based on spatio-temporal information embedding. Then, we design an architecture based on discrete wavelet decomposition, which uses denoising diffusion network and spatio-temporal neural network to process trend components and fluctuation components, respectively. A series of experimental results on a variety of real-world datasets represent that our method achieves the state-of-the-art probabilistic prediction performance.