ML-Driven Energy Savings for Cellular Baseband Units via Traffic Prediction

Abstract

5G networks continue to expand as connected devices and traffic surge worldwide. This growth presents an opportune moment to address the ongoing challenge of Baseband Unit (BBU) energy consumption in large-scale deployments. Traditional static energy management approaches frequently waste resources and lead to increased costs, highlighting the need for more dynamic methods that adapt to changing network conditions. This paper introduces the Predictive Energy Saver for Baseband Units (PESBiU) 2.0, a new framework designed to address this challenge based on the premise that a balanced and advanced combination of precise traffic prediction and intelligent power-state decision-making can achieve superior energy savings without compromising user Quality of Service (QoS). PESBiU 2.0 uses granular interval datasets and machine learning (ML) models to predict traffic loads and optimize power states. The design features a hybrid architecture of Hyper Convolutional Neural Network-Long Short-Term Memory (Hyper-CNN-LSTM) model for accurate forecasting with a reinforcement learning (RL) decision engine based on Dueling Double Deep Q-Networks (DDDQN), making it the first framework to apply DDDQN for BBU energy optimization in 5G and beyond networks. Evaluation results confirm that PESBiU 2.0 effectively balances complexity and performance, achieving more than 40% reduction in BBU power consumption without compromising service quality. This benefits operators, researchers, and vendors seeking improved energy efficiency and consistent performance in 5G+ networks. The findings indicate a clear path for integrating advanced ML methods to enhance network efficiency and reliability, offering a scalable solution for future telecommunications.