Capturing Dynamic Dependencies and Temporal Fluctuations for Traffic Flow Forecasting

Urban transportation efficiency and reliability are crucial. Most approaches use predefined or adaptive graphs to model spatio-temporal dependencies, but they struggle to make accurate predictions due to the dynamic nature of traffic conditions, potential data drift, and inaccuracies in spatio-temporal patterns. In addition, it is difficult for existing techniques to resolve the complex dependencies between spatial and temporal dynamics when traffic data exhibits fluctuations in response to temporal patterns. To address these issues, we propose the CDDTF model, which uses a novel approach with learnable traffic embeddings and graph convolutions in its encoder, effectively adapting to real-time changes in traffic conditions. The decoder uses Temporal Interleaved Convolution and a Dual-Branch Gated Attention to combat Temporal Fluctuations during peak hours. Comparative experiments on four real-world datasets show that CDDTF outperforms contemporary state-of-the-art methods in terms of prediction accuracy.