A Data-Driven Framework for Energy Consumption Forecasting in Disaggregated O-RAN Architectures

The evolution of cellular networks has reached an unprecedented volume of users, who demand high-quality service. The energy consumption from the networks has grown alongside the number of users, while the requirement for energy sustainability is more constraining. The O-RAN paradigm comes with a flexible and operator-collaborative vision for the infrastructure of cellular networks. This approach adopts a disaggregated modularization of the radio access, allowing for more controlled energy management in each network element. This work explores the energy forecasting possibilities within a real O-RAN infrastructure, from which real energy consumption data is obtained and a forecasting algorithm is evaluated. The evolution of cellular networks has reached an unprecedented volume of users, who demand high-quality service. The energy consumption from the networks has grown alongside the number of users, while the requirement for sustainability is becoming increasingly constrained. The O-RAN paradigm offers a flexible and operator-collaborative vision for cellular network infrastructures. This adopts a disaggregated modularization of the radio access (into CU, DU and RU), enabling more controlled and dynamic energy management in each network element. Leveraging this potential, the work presented in this paper explores the potential of power consumption forecasting within a real O-RAN infrastructure. Apart from the self-forecasting within network elements, the existing equipment hierarchy and relationship among network parameters expand the possibilities to forecast from cross-elements and cross-parameters perspectives. For this purpose, statistical and machine learning-based algorithms are evaluated with real data, exhibiting low error metrics in terms of RMSE (<0.3W ), MAE (<0.2W ) and MAPE (< 0.17%) for the noisiest element (CU). Moreover, by means of the MAPIE technique, error margins are calculated as bounds for the forecast.