Multi-time Scale Augmented Neural ODEs graph neural for traffic flow prediction with elastic channel variation

Abstract

Traffic flow prediction is a critical task in traffic management due to the complex and dynamic spatio-temporal correlations inherent in traffic data. While existing methods often employ graph convolutional networks (GNNs) and temporal extraction modules to model spatial and temporal dependencies, respectively, deep GNNs suffer from oversmoothing, which impairs their ability to capture long-term spatial relationships. Augmented Neural Ordinary Differential Equations (ANODEs) offer a solution to this issue by enabling deeper models without oversmoothing, but they struggle with the complexity and variability of traffic data, leading to poor prediction performance. In this study, we propose the Multi-time Scale Augmented Neural ODEs Graph Neural Network (MTEC-AODE) for Traffic Flow Prediction. To address the challenges of complex information processing, we introduce the Elastic Channel Variation strategy, which adjusts the number of channels dynamically. Furthermore, we construct a traffic semantic neighborhood matrix using a Gaussian kernel similarity matrix, which captures semantic relationships across regions, aiding in the construction of a global dynamic traffic model. To handle the variability of traffic data, we develop the Multi-time Scale Augmented Neural ODEs Solver, allowing the model to adapt to different time scales and respond to dynamic changes in traffic patterns. We evaluate our model on several real-world traffic datasets, achieving Mean Absolute Errors (MAE) of 2.01, 3.52, 2.96, 3.20, 15.59, 19.75, 22.14, and 16.22, and Mean Absolute Percentage Errors (MAPE) of 4.48%, 10.17%, 7.22%, 7.66%, 15.21%, 13.98%, 9.72%, and 10.2%. Experimental results show that our method outperformed state-of-the-art benchmarks.