ST-VGBiGRU: A Hybrid Model for Traffic Flow Prediction With Spatio-Temporal Multimodality

Yin, Lisheng; Liu, Pan; Wu, Yangyang; Shi, Cheng; Wei, Xinyue; He, Yigang · 2023 · DOAJ

DOI: 10.1109/ACCESS.2023.3282323

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of accurate traffic flow prediction, a critical component for intelligent transportation systems. Traffic data exhibits complex short-term non-stationarity, spatial correlation, and temporal correlation, which traditional statistical methods and standard deep learning models struggle to capture effectively. Existing approaches often fail to handle the non-stationary nature of traffic flows at fine time scales or rely on Euclidean grid conversions that lose spatial topology information. To overcome these limitations, the authors propose ST-VGBiGRU, a hybrid spatio-temporal prediction model that integrates improved Variational Mode Decomposition (VMD), Graph Attention Networks (GAT), and Bidirectional Gated Recurrent Units (BiGRU). The methodology involves a three-stage processing pipeline. First, an improved VMD module decomposes the non-stationary traffic flow sequence into relatively stationary modal components. To enhance decomposition accuracy, the authors introduce Fuzzy Entropy to distinguish between low-frequency components and high-frequency noisy components, applying threshold noise reduction to the latter. Second, a GAT module captures spatial correlations by dynamically calculating attention coefficients between a central prediction node and its neighboring nodes, effectively handling the non-Euclidean structure of road networks. Third, each modal component, now enriched with spatial features, is fed into a BiGRU network to extract temporal correlations. The model parameters are optimized using an improved RMSProp algorithm to accelerate convergence and improve prediction accuracy. The study evaluates the ST-VGBiGRU model using the RTMC and PeMS traffic datasets. Ablation experiments on the RTMC dataset demonstrate that each component—the improved VMD, GAT, and BiGRU modules—positively contributes to overall performance. Comparative baseline experiments against six other models on both datasets show that ST-VGBiGRU outperforms existing methods in prediction accuracy. The results indicate that the hybrid approach effectively mitigates the impact of short-term non-stationarity while robustly capturing spatio-temporal dependencies. The significance of this work lies in its comprehensive handling of traffic flow characteristics. By combining signal decomposition with graph-based spatial attention and bidirectional temporal learning, the model provides a more accurate and efficient solution for traffic prediction. The integration of Fuzzy Entropy for noise reduction and the use of GAT for spatial feature extraction address specific weaknesses in prior models, such as modal mixing in decomposition and rigid spatial assumptions in convolutional networks. This approach offers a robust framework for intelligent traffic management, enabling better anticipation of road conditions and alleviation of congestion.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• mental model of traffic