DOI: 10.65520/erciyesfen.1968962 ISSN: 1012-2354

Energy Consumption Prediction for 5G Base Stations: A Comparative Machine Learning Study under Hardware Heterogeneity

Ramazan Enisoğlu, Burak Alp İnan
Show PDF Cite
Background/Aim: Energy consumption of 5G base stations has become a central operational concern, and energy-saving mechanisms such as carrier shutdown and sleep modes require accurate prediction of station-level consumption. This study compares prediction models under conditions reflecting deployment to unseen base stations.Methods: Seven models – from a trivial baseline and linear models to tree ensembles and a multilayer perceptron – were evaluated on the ITU 5G Base Station Energy Consumption Modelling dataset (98,084 hourly observations from 923 base stations with heterogeneous hardware) using station-grouped five-fold cross-validation, station-clustered bootstrap intervals, repeated station partitions and out-of-time tests. An analytical load-dependent power model and four published data-driven approaches were re-implemented under the identical protocol.Results: LightGBM attained the best accuracy among the seven models (R² = 0.910, WMAPE = 10.22%), matched in R² by the multilayer perceptron at roughly twenty times the training cost. LightGBM remained within 0.15 percentage points of the most accurate re-implemented technique, whereas the analytical model was clearly less accurate; lower published errors (WMAPE of about 4–5%) arise only when test stations were seen during training. Consumption was governed predominantly by static hardware attributes and their interaction with traffic load. Accuracy was stable across the observed days, but tree ensembles underestimated consumption when load exceeded the training range.Conclusion: Gradient boosting offers the most favourable accuracy–cost trade-off for unseen base stations, while seasonal use requires periodic recalibration. The study also documents a sparsity-induced instability in linear models and the hardware distribution shift between training and evaluation data.

More from our Archive