A framework for traffic flow forecasting using graph attention networks with external factor integration

Abstract

External factors against traffic patterns are one of the most important parameters that influences the prediction of traffic flow rates. The study aimed to offer advancement over Graph Convolutional Networks (GCNs) by overcoming the limitation of GCNs, which depend on node degrees to assess importance. The study proposes traffic forecasting framework using Graph Attention Networks (GATs), incorporating external factors such as weather and public events, which significantly influence traffic patterns. Recent advancements in machine learning, particularly deep learning and graph neural networks, have significantly improved traffic forecasting accuracy. However, these models often neglect these external factors. The framework using deep learning techniques, with a focus on Graph Attention Networks (GATs) integrates spatial, temporal and contextual data. This led to achieving forecasting accuracy with superior performance in terms of metrics like Mean Squared Error (MSE), Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). The study achieved 7.017 × 10⁻7 MSE, 0.0010RMSE and 0.176 MAPE. Experimental results demonstrate the effectiveness of the proposed approach in handling real-world traffic complexities. In addition, the study evaluates the effect of external factors on traffic predictions, and the findings emphasize the importance of incorporating contextual data in traffic forecasting to enhance prediction reliability. Comparative analysis shows that the proposed GAT-based framework outperforms existing methods, including STGCN and GRAM-ODE.  However, this reveals new challenges and research opportunities for future studies.