A Wavelet Disentanglement and topological semantic neural network for traffic flow forecasting

Existing spatio-temporal models struggle to capture spatial dependencies in traffic flow that involve both local topological structures and semantic relations shaped by multiple factors. To this end, a novel Wavelet disentanglement and Topological Semantic Neural Network (WTSNet) for accurate traffic flow forecasting is presented. We employ Discrete Wavelet Transform (DWT) to decouple traffic flow into stable trends and event fluctuations. To separately model the stable trends and event fluctuations, the dual-channel spatial semantic layer is designed. It can effectively handle traffic flow affected by multiple modalities and solve the limitations of traditional models in capturing global semantic dependencies. To better capture both spatial topological dependencies and temporal dependencies the spatial topological and temporal layer is also designed in WTSNet. The designed layer can enhance the model's ability to learn both local topological dependencies and global semantic features in traffic flow. Extensive experimental results on four real public transportation datasets show that the proposed approach has better performance.