Dynamic Voltage Optimization for Energy Efficient Radios

Abstract

Reducing energy waste in radios is essential for lowering the CO2 footprint of cellular connectivity. In current systems, radios operate at a static high voltage regardless of their physical resource block utilization (PRB-U), resulting in inefficiencies. This thesis proposes a machine learning (ML)-based approach to enable dynamic voltage control and evaluates which models and task formulations are most suitable for this purpose. Eight ML models were tested across two formulations: (1) classification of voltage levels and (2) regression to predict PRB-U values and map them to voltage levels.

The results show that ML can significantly reduce energy waste in certain radios, although effectiveness varies by device - some perform comparably using simpler methods. Classification outperformed regression in reducing voltage underestimations.

The Random Forest (RF) classifier and a customized Feedforward Neural Network (FFNN) classifier emerged as top performers. The FFNN achieved an F1 score of 0.35, saved 12.41% energy, and resulted in 34 underestimations. The RF reached an F1 score of 0.29, saved 12.43%, and had 42 underestimations. In contrast, the bestperforming baseline model had an F1 score of only 0.19 with 55 underestimations out of 68, underscoring the benefit of ML-based approaches.

Shifting the classification threshold helped manage the trade-off between energy savings and underestimations. Notably, the FFNN achieved just 5 underestimations while maintaining 10% energy savings.

SHAP-based explainability analysis showed that PRB-U at the current timestep was the most influential feature. The RF leveraged a broader feature set, while the FFNN focused more narrowly on a dominant input.

In conclusion, this thesis demonstrates that intelligent, ML-driven voltage control can enhance energy efficiency in radios. It also emphasizes the importance of balancing energy savings and prediction errors during model deployment, potentially by adjusting class probability thresholds.