A Novel Traffic Prediction Method Using Machine Learning for Energy Efficiency in Service Provider Networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Energy Systems

Background:

• Data centers face increasing energy demands, necessitating efficient operational strategies.
• Accurate prediction of energy consumption is crucial for optimizing resource allocation and reducing environmental impact.
• Traditional prediction methods may not adequately capture the dynamic nature of energy usage in complex systems.

Purpose of the Study:

• To present a systematic methodology for complex prediction problems, emphasizing energy efficiency.
• To evaluate the performance of various recurrent and sequential neural networks for data center energy prediction.
• To identify the most accurate and computationally efficient neural network model for energy efficiency applications.

Main Methods:

• Utilized recurrent and sequential neural networks for prediction tasks.
• Conducted a case study in the telecommunications industry focusing on data center energy efficiency.
• Compared four neural network models: Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRUs), and Online Sequential Extreme Learning Machine (OS-ELM).
• Evaluated models based on prediction accuracy and computational time using real traffic data.

Main Results:

• The Online Sequential Extreme Learning Machine (OS-ELM) model achieved superior prediction accuracy compared to RNNs, LSTM, and GRUs.
• OS-ELM also demonstrated higher computational efficiency, indicating faster processing times.
• Simulations using real traffic data projected potential daily energy savings of up to 12.2%.

Conclusions:

• The proposed systematic approach, particularly using OS-ELM, is highly effective for energy efficiency prediction in data centers.
• The methodology shows significant potential for reducing energy consumption and can be adapted to other industries.
• Further advancements in technology and data will enhance the applicability of this approach for diverse prediction challenges.