Spatial-Temporal Traffic Flow Prediction With Fusion Graph Convolution Network and Enhanced Gated Recurrent Units

• Published in: IEEE access 2024, Vol.12, p.56477-56491
• Main Authors: Cai, Chuang, Qu, Zhijian, Ma, Liqun, Yu, Lianfei, Liu, Wenbo, Ren, Chongguang
• Format: Article
• Language: English
• Subjects: Adaptation models; Artificial neural networks; Convolution; Convolutional neural networks; Correlation; Datasets; Feature extraction; gated recurrent unit; graph convolution network; Logic gates; Performance prediction; Prediction models; Predictive models; Recurrent neural networks; residual connect; Roads; Root-mean-square errors; spatio-temporal correlation; Spatiotemporal phenomena; System effectiveness; Traffic control; Traffic flow; Traffic flow prediction; Traffic models; Transportation networks; Transportation systems
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3349690

Accurately predicting traffic flow is paramount for the efficient operation of transportation systems. The key to enhancing prediction accuracy lies in effectively mining the intricate spatio-temporal correlations within traffic flow data. However, traditional traffic flow prediction methods that combine Graph Convolutional Network and Recurrent Neural Network have limitations in capturing comprehensive spatial correlation information and face challenges in modeling long-term temporal dependencies, consequently leading to suboptimal prediction performance. This study proposes a hybrid traffic flow prediction model based on fusion graph convolutional network and enhanced gate recurrent unit. Initially, a fusion graph structure is constructed based on adjacency graph and adaptive graph to better represent the correlations between nodes in the road network. Subsequently, the stacked fusion graph convolution module is utilized to capture multi-level spatial correlations and the enhanced gated recurrent unit is applied to extract multi-scale temporal correlations. In addition, the model integrates the extracted spatio-temporal features with the direct features through residual connection units, and utilizes the fused features for prediction, achieving superior predictive performance. The experimental results from four authentic datasets demonstrate that our proposed model outperforms state-of-the-art baseline models, showcasing an average enhancement of 3% in Mean Absolute Error(MAE), 3.3% in Root Mean Square Error(RMSE), and 2.7% in Mean Absolute Percentage Error(MAPE) across the four datasets.