AI for Link Adaptation and Energy Prediction in Realistic 5G Networks

Sammanfattning

As 5G networks grow in complexity, ensuring efficient spectrum use and energy consumption has become a critical challenge, one where AI offers significant promise. This paper explores AI-driven optimization of link adaptation in 5G standalone networks, focusing on MCS prediction using SINR and CQI as input. With more than 20,000 samples generated through 5G-LENA simulations, several supervised models, including CNN, LSTM, and contextual variants, were evaluated. While overall classification accuracy remained modest (52–55%), recurrent models outperformed others, aligning with prior research on temporal dependencies in wireless channels. Parallel efforts in energy modeling used a public dataset of 5G base station activity. Traditional regressors struggled with high error, but a neural model inspired by (Chen et al., 2024) achieved a MAPE of 5.5%, underscoring the importance of temporal and contextual features in forecasting energy use. Together, these results highlight the practical potential of AI in 5G network optimization while emphasizing the need for richer datasets and more temporally expressive models.