When Time Series Data Become Images: U-Net for Network Traffic Prediction

Accurate network traffic prediction is a fundamental requirement for proactive network management, enabling efficient resource provisioning, congestion avoidance, and quality-of-service (QoS) assurance. While recent deep learning approaches have shown strong capability in modeling spatio-temporal traffic dynamics, many existing methods rely on recurrent or graph-based architectures that present high computational overhead and limited scalability. In this paper, we propose an efficient U-Net-based framework for network traffic prediction that reformulates traffic forecasting as a spatio-temporal image-to-image regression problem. By representing historical traffic states as two-dimensional traffic maps, the proposed model exploits the encoder-decoder architecture and skip connections of UNet to jointly capture global traffic trends and fine-grained local variations. Extensive experiments on a real-world network traffic dataset show that the proposed approach consistently outperforms representative state-of-the-art models, achieving up to approximately 80% reduction in prediction error compared to state-of-the-art adversarial-based and its variant baselines across multiple forecasting horizons. These results indicate that fully convolutional U-Net architectures provide an effective and scalable solution for accurate network traffic prediction.