Traffic Flow Prediction Using Spatio-Temporal Graph Neural Networks Based on Hybrid Adaptive Feature Enhancement

Aiming at the poor adaptation of most current models to dynamic maps and the difficulty in capturing long-term spatio-temporally relevant features, this paper proposes a hybrid adaptive feature-enhanced graph neural network (HAFEGNN) model. To address these challenges, we propose a new approach based on graph neural networks that integrates multiple state-of-the-art enhancement feature representation mechanisms. The model enhances the representation of dynamic spatio-temporal features by integrating multi-scale convolution, channel self-attention and spatial self-attention mechanisms. On top of that, temporal self-attention mechanism is also introduced to focus on learning long-term dependencies, which leads to better understanding of patterns in historical data and making more accurate predictions. The model further incorporates Dynamic Graph Convolutional Recurrent Network (DGCRN) to capture spatio-temporal dynamics. To enhance the adaptability to dynamic graph structures, it integrates spatial location embedding with graph attention mechanisms to form a spatially aware module, which improves the prediction accuracy. Our experimental evaluation on the widely used PeMS04 and PeMS08 datasets demonstrates the effectiveness of the proposed HAFEGNN. The model achieves significant improvements in key metrics such as MAE, RMSE, and MAPE, outperforming existing state-of-the-art methods and validating its ability to handle complex traffic data. This study provides a new solution for traffic flow prediction and demonstrates its potential in handling complex traffic data.