Dynamic Learning Framework for Smooth-Aided Machine-Learning-Based Backbone Traffic Forecasts.

Area of Science:

• Network engineering
• Data science
• Telecommunications

Background:

• Increasing demand for big data, IoT, 5G, and V2X services necessitates precise network resource planning.
• Maintaining Quality of Service (QoS) requires accurate forecasting for effective resource allocation.

Purpose of the Study:

• To propose a reliable hybrid dynamic bandwidth slice forecasting framework.
• To enhance network forecasting models by integrating Long Short-Term Memory (LSTM) neural networks with local smoothing techniques.
• To develop a framework that dynamically adapts to changes in data series.

Main Methods:

• A hybrid framework combining LSTM neural networks and local smoothing methods.
• Validation using backbone traffic data.
• Implementation of hybrid moving average LSTM (MLSTM) and robust locally weighted scatter plot smoothing and LSTM (RLWLSTM) models.

Main Results:

• Significant improvements in forecasting accuracy were observed.
• MLSTM achieved 28% and 24% improvement for LTE, and 35% and 32% for MPLS time series.
• RLWLSTM showed a 45% improvement for upstream traffic.
• The dynamic learning framework demonstrated up to 100% improvement.

Conclusions:

• The proposed hybrid framework effectively improves network bandwidth slice forecasting accuracy.
• The integration of LSTM with local smoothing methods offers a robust solution for dynamic network environments.
• The framework minimizes data loss during the smoothing process, ensuring reliable forecasts.