Research on Satellite Network Traffic Prediction Based on Improved GRU Neural Network

Area of Science:

• Computer Science
• Telecommunications Engineering
• Artificial Intelligence

Background:

• Current satellite network traffic forecasting methods struggle with long-term correlations in traffic data, leading to significant errors and reduced accuracy.
• Existing models often fail to fully exploit the self-similarity and long-range dependencies inherent in satellite traffic sequences.
• This limitation necessitates advanced methods for more precise and reliable network traffic prediction.

Purpose of the Study:

• To develop an improved satellite network traffic forecasting method that addresses the limitations of existing techniques.
• To enhance prediction accuracy by effectively utilizing the long correlation characteristics of satellite traffic data.
• To improve forecasting efficiency through optimized model hyperparameter tuning.

Main Methods:

• Proposes an improved Gate Recurrent Unit (GRU) neural network model integrated with an attention mechanism.
• The attention mechanism focuses on the importance of traffic data and hidden states, capturing time-dependent and interdependent characteristics.
• Employs the Particle Swarm Optimization (PSO) algorithm to determine optimal hyperparameters for the GRU model, enhancing prediction efficiency.

Main Results:

• The proposed attention-based GRU model demonstrates superior fitting with real satellite traffic data compared to traditional methods.
• Achieved significant error reductions: 26.9% compared to GRU, 37.2% compared to Support Vector Machine (SVM), and 57.8% compared to Fractional Autoregressive Integration Moving Average (FARIMA).
• The integration of PSO effectively optimized model hyperparameters, leading to improved prediction efficiency.

Conclusions:

• The enhanced GRU model effectively mines self-similarity and long-term correlations in satellite traffic data for improved forecasting.
• The attention mechanism and PSO optimization contribute to higher prediction accuracy and efficiency in satellite network traffic forecasting.
• This method offers a promising solution for reducing errors and enhancing the reliability of satellite network traffic predictions.