A novel hybrid framework based on temporal convolution network and transformer for network traffic prediction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Telecommunications Engineering

Background:

• Accurate mobile network traffic prediction is crucial for rational resource allocation and ensuring stable, fast network services.
• The inherent burstiness and uncertainty of network traffic present significant challenges to precise prediction.
• Understanding spatio-temporal correlations is key to improving traffic forecasting models.

Purpose of the Study:

• To develop a novel deep-learning model for accurate mobile network traffic time-series prediction.
• To effectively extract and utilize spatio-temporal features from network traffic data.
• To enhance the efficiency and accuracy of mobile network resource management.

Main Methods:

• Proposed a deep-learning model integrating Convolutional Block Attention Module (CBAM) Spatio-Temporal Convolution Network (CSTCN) and Transformer with a sparse self-attention mechanism.
• Utilized an improved TCN with CBAM for spatial feature extraction, forming the CSTCN component.
• Employed Transformer with sparse self-attention to further capture complex spatio-temporal dependencies.

Main Results:

• The CSTCN-Transformer model demonstrated significant improvements in prediction accuracy compared to baseline models.
• Achieved substantial reductions in mean square error (up to 65.16%) and mean average error (up to 53.10%) on a real-world dataset.
• Validated the model's effectiveness on a mobile network traffic dataset from Milan.

Conclusions:

• The proposed CSTCN-Transformer model effectively addresses the challenges of mobile network traffic prediction.
• The integration of CBAM, TCN, and sparse self-attention Transformer significantly enhances spatio-temporal feature extraction and prediction accuracy.
• The model offers a promising solution for optimizing mobile network operations and user experience.