A Traffic Flow Prediction Model with Dynamic Graph Convolutional Neural Network Based on Generative Adversarial Learning

Deep learning has become a central methodology in predicting traffic flow at the road network scale. Existing methods relying on mean absolute error as the loss function often result in blurred predictions, hindering accuracy at the node level. To address this, this paper introduces Multi-time-scale Wasserstein Generative Adversarial Network (MT-WGAN), which employs STS-DGCN as generator for capturing complex spatio-temporal relationships within traffic data. A fully connected network serves as the discriminator, engaging in adversarial learning with the generator. The model leverages the Wasserstein distance to quantify disparities in data distributions between predicted and observed values, and improves prediction accuracy by approximating the true distribution of traffic flow data. Experiments confirm that the proposed approach exceeds the performance of other advanced baseline methods on public datasets, delivering superior prediction accuracy.