Abstract
Accurate and robust traffic flow prediction remains a fundamental yet unresolved challenge in intelligent transportation systems, primarily due to the intertwined nature of long-term trends and short-term disruptions, alongside complex, dynamic spatio-temporal dependencies. This study proposes the Disentangled Spatio-Temporal Graph Attention Mamba (DSTGA-Mamba) model, a unified framework that comprises the following key components. First, we introduce a multi-level wavelet-based disentanglement module to explicitly decouple temporal dynamics into trend and event components, reducing distribution shifts and long-horizon error accumulation. Second, we design a dual-channel spatio-temporal encoder combining global temporal attention and causal convolution, integrated with a novel efficient spectral graph attention mechanism to capture scalable and expressive cross-node interactions. Third, we embed a selective state-space model (ST-Mamba) to capture ultra-long temporal dependencies with linear complexity, enhancing both scalability and forecasting fidelity. Complemented by a multi-supervision decoder that fuses trend-guided supervision with adaptive event modeling, DSTGA-Mamba demonstrates superior performance across four benchmark datasets, achieving consistent gains in accuracy, robustness, and computational efficiency over strong baselines.