A temporal decoupling wavelet graph network with direction-aware diffusion for on-ramp flow prediction

Accurate on-ramp flow prediction is essential for proactive ramp metering and freeway efficiency optimization. However, on-ramp flows are characterized by small magnitudes, strong volatility, and heterogeneous temporal patterns across multiple frequency scales, which are further compounded by periodic temporal patterns associated with calendar information. To address these challenges, this paper proposes the Temporal Decoupling Wavelet Graph Network (TDWGNet). The model first employs a discrete wavelet transform to decompose on-ramp flow into one trend and multiple detail components. A shared LSTM arrangement encodes these components, after which a temporal pattern decoupling module reduces the influence of periodic calendar-related patterns in the trend representation through orthogonal projection. The refined trend and aggregated detail features are fused asymmetrically, preserving the dominant role of the trend while incorporating corrective detail information. A directed diffusion graph convolution, built upon physical route distances and a direction mask, captures the asymmetric spatial propagation along the freeway. Extensive experiments on real-world on-ramp datasets and PeMS dataset demonstrate that TDWGNet consistently outperforms 7 baseline models spanning recurrent architectures, attention-based methods, graph convolutional networks, and recent decoupling-based approaches, achieving consistently strong performance across the three on-ramp datasets and competitive results on the PeMS dataset. Ablation studies confirm that the complete model generally surpasses its ablated variants over longer prediction windows, validating the complementary roles of wavelet decomposition, temporal pattern decoupling, and directed spatial encoding. Parameter sensitivity analysis further reveals that dataset-specific tuning can yield additional improvements, thereby demonstrating the effectiveness and stability of the proposed architecture.