Deep Learning Based Traffic Prediction Method for Digital Twin Network

Network traffic prediction (NTP) can predict future traffic leveraging historical data, which serves as proactive methods for network resource planning, allocation, and management. Besides, NTP can also be applied for load generation in simulated and emulated as well as digital twin networks (DTNs). This paper focuses on accurately predicting background traffic matrix (TM) of typical local area network (LAN) for traffic synchronization in DTN. A survey is firstly conducted on DTN, conventional model, and deep learning based NTP methods. Then, as the major contribution, a linear feature enhanced convolutional long short-term memory (ConvLSTM) model based NTP method is proposed for LAN. An autoregressive unit is integrated into the ConvLSTM model to improve its linear prediction ability. In addition, this paper further optimizes the proposed model from both spatial and channel-wise dimensions. Particularly, a traffic pattern attention (TPA) block and a squeeze & excitation (SE) block are derived and added to the enhanced ConvLSTM (eConvLSTM) model. Comparative experiments demonstrate that the eConvLSTM model outperforms all the baselines. It can improve the prediction accuracy by reducing the mean square error (MSE) up to 10.6% for one-hop prediction and 16.8% for multi-hops prediction, compared to the legacy CovnLSTM model, with still satisfying the efficiency requirements. The further enhancement of the eConvLSTM model can additionally reduce the MSE about 2.1% for one-hop prediction and 4.2% for multi-hops prediction, with slightly degrading efficiency. The proposed eConvLSTM model based NTP method can play a vital role on DTN traffic synchronization.