U-GRU for Short-Term Urban Traffic Speed Forecasting with Ordered-Node Feature Transformation and Channel–Temporal Recalibration
TL;DR
This study proposes U-GRU, a staged forecasting model that integrates node-wise gated recurrent temporal encoding, a one-dimensional U-Net-based ordered-node feature transformation module, external-feature alignment and channel–temporal recalibration for short-term traffic speed forecasting.
Abstract
Short-term traffic speed forecasting based on sensor-derived traffic signals is important for route guidance, adaptive signal control, congestion management and intelligent transportation services. Accurate prediction remains challenging because urban traffic states are influenced by nonlinear temporal dynamics, inter-node relationships, heterogeneous contextual information and uncertainty in external conditions. This study proposes U-GRU, a staged forecasting model that integrates node-wise gated recurrent temporal encoding, a one-dimensional U-Net-based ordered-node feature transformation module, external-feature alignment and channel–temporal recalibration. Unlike graph-convolutional methods that explicitly propagate information through a predefined or learned road topology, U-GRU investigates whether a transformation path along the stored node order in the data files can provide a useful representation under an ordered-node assumption. POI-related and weather-related auxiliary matrices are used as external representations provided in the public dataset; because field-level semantic metadata are unavailable, they are treated as dataset-level auxiliary inputs rather than manually interpreted variables. The experiments use chronological train/validation/test splitting, validation-based checkpoint selection, five random seeds for U-GRU, an expanded baseline set including the STGformer baseline evaluated under the same SZ-taxi setting, ablation studies, weather-validity analysis and node-order sensitivity analysis. On the SZ-taxi dataset, U-GRU reports MAE values of 2.70–2.72, RMSE values of 4.04–4.12 and R2 values of 0.83–0.85 across the 15–60 min forecasting horizons. The results suggest comparable or slightly better short-term forecasting performance under the evaluated dataset and protocol, while the conclusions remain dataset-specific and should be further examined on additional traffic networks.