Prediction of telephone calls load using Echo State Network with exogenous variables

Area of Science:

• Telecommunications Engineering
• Machine Learning
• Time Series Analysis

Background:

• Accurate forecasting of incoming call loads is crucial for efficient mobile network resource management.
• Traditional forecasting methods may not fully capture the complex dynamics of cellular network traffic.

Purpose of the Study:

• To improve mobile network call load forecasting accuracy.
• To investigate the utility of exogenous variables in Echo State Network (ESN) models.
• To explore novel training methodologies and parameter optimization techniques.

Main Methods:

• Utilized Echo State Networks (ESNs) for call load forecasting.
• Incorporated exogenous telephone activity records as external input variables.
• Analyzed novel readout training methods: ν-SVR and elastic net penalty.
• Employed genetic algorithms for parameter tuning and feature selection.

Main Results:

• Demonstrated the effectiveness of incorporating exogenous variables in ESN forecasting.
• Identified optimal parameter settings and informative external time-series through genetic algorithms.
• Achieved competitive or superior performance compared to standard prediction models.

Conclusions:

• Echo State Networks, augmented with exogenous data and advanced training/optimization, offer a powerful approach for mobile network call load forecasting.
• The proposed methods provide a robust framework for enhancing the accuracy and efficiency of telecommunications traffic prediction.