WGM-Net: A multi-wavelet and geography-aware dynamic memory network with temporal alignment Mamba for traffic flow forecasting

Traffic flow forecasting is a critical task in intelligent transportation systems, where accurate prediction requires modeling multi-scale temporal dynamics, regional functional heterogeneity, and distribution shifts from historical observations to future prediction targets. Existing methods often process mixed-frequency traffic signals directly or rely on a single wavelet basis for decomposition, which limits their ability to distinguish stable trends from abrupt fluctuations. In addition, geographic and calendar-related priors are usually incorporated through shallow feature fusion, making it difficult to capture their interactions with evolving traffic patterns. To address these issues, this paper proposes WGM-Net, a multi-wavelet and geography-aware dynamic memory network for traffic flow forecasting. The proposed model uses complementary wavelet bases to decompose traffic-flow variations into trend and detail representations, thereby enhancing multi-scale temporal feature extraction. It further incorporates holiday-aware temporal embeddings and Point-of-Interest-driven geographic representations into a dynamic memory structure to capture calendar-specific variations and regional functional differences. Moreover, a temporal alignment Mamba mechanism alleviates history-target distribution shifts and refines the prediction sequence, contributing to improved multi-step forecasting performance. Extensive experiments on two real-world traffic datasets, PEMSD4 and PEMSD8, demonstrate that WGM-Net consistently outperforms competitive baseline models in forecasting accuracy. These results indicate that WGM-Net provides an effective and robust framework for modeling complex spatio-temporal traffic patterns in urban road networks. The source code for WGM-Net is publicly available at: https://github.com/Linly01/WGM-Net .