Deep Learning Based Traffic Prediction Method for Digital Twin Network.

Area of Science:

• Computer Science
• Network Engineering
• Artificial Intelligence

Background:

• Network traffic prediction (NTP) is crucial for network management and resource allocation.
• Digital Twin Networks (DTNs) require accurate traffic synchronization for effective simulation and emulation.
• Existing NTP methods have limitations in predicting complex traffic patterns.

Purpose of the Study:

• To propose an accurate NTP method for local area networks (LANs) to support traffic synchronization in DTNs.
• To enhance deep learning models for improved traffic matrix (TM) prediction.
• To evaluate the performance of the proposed model against baseline methods.

Main Methods:

• A survey of existing DTN, conventional, and deep learning-based NTP methods.
• Development of a linear feature enhanced Convolutional Long Short-Term Memory (ConvLSTM) model.
• Integration of an autoregressive unit for linear prediction enhancement.
• Optimization using Traffic Pattern Attention (TPA) and Squeeze & Excitation (SE) blocks, creating the eConvLSTM model.

Main Results:

• The eConvLSTM model significantly outperforms baseline methods in NTP accuracy.
• Reduced Mean Square Error (MSE) by up to 10.6% (one-hop) and 16.8% (multi-hop) compared to the legacy ConvLSTM.
• Further enhancements to eConvLSTM reduced MSE by an additional 2.1% (one-hop) and 4.2% (multi-hop).
• The model satisfies efficiency requirements for practical applications.

Conclusions:

• The proposed eConvLSTM model offers a superior approach to NTP for LANs.
• This method is vital for achieving accurate traffic synchronization in DTNs.
• The model provides a robust and efficient solution for network resource management and simulation.