Energieffektiv AI/ML för prognoser i Radio Access Networks

In the development of artificial intelligence models, the primary focus is often on maximizing predictive performance, with little regard for the energy or computational costs involved. As artificial intelligence models are increasingly deployed at scale, their energy demands have become a growing concern. In telecom networks, where energy forecasting models could be deployed across thousands of base stations, scalability and efficiency are essential considerations. This thesis investigates how machine learning can be used to forecast radio energy consumption in an energy-aware manner, aiming to identify models that perform well while minimizing energy usage during both training and inference. Several machine learning models were selected and evaluated using real 4G telecom data. The chosen models, SparseTSF, HADL, LTBoost, and a Green Accelerated Hoeffding Tree, were implemented and tested on multiple forecast horizons. Forecasting performance was assessed using mean squared error, mean absolute error, and relative squared error. Energy consumption during both training and inference phases was measured using Kepler, a power monitoring tool designed for containerized environments. The results showed that SparseTSF and HADL achieved the best forecasting accuracy while maintaining low training energy use. GAHT was the most efficient during inference, and LTBoost provided a balance across all evaluated metrics. These outcomes highlight that lightweight neural networks and tree-based models are strong candidates for energy-constrained environments. Additionally, optimization techniques such as feature reduction, sparse training, quantization, and energy-optimized early stopping were applied to each model. Feature reduction consistently lowered energy use with minimal accuracy loss, while sparse training and energy-optimized early stopping proved to be more situational. This thesis demonstrates that improving energy efficiency in forecasting tasks is achievable without severely compromising accuracy. Careful model selection and targeted optimizations offer practical paths toward more sustainable deployment of machine learning in time series forecasting or telecom settings. The results show that energy-efficient models did not perform worse than models with higher energy consumption, suggesting that strong forecasting performance can be achieved without resorting to more resource-intensive architectures, and that integrating energy considerations early in the development process leads to more efficient machine learning systems.