The Open Radio Access Network (RAN) paradigm envisions a more flexible, interoperable, and intelligent RAN ecosystem via new open interfaces and elements like the RAN Intelligent Controller (RIC). However, the impact of these elements on Open RAN's power consumption remains heavily unexplored. This work for the first time evaluates the impact of Key Performance Indicator (KPI) monitoring on RIC's power consumption using real traffic and power measurements. By analyzing various RIC-RAN communication scenarios, we identify that RIC's power consumption can become a scalability bottleneck, particularly in large-scale deployments, even when RIC is limited to its core operational functionalities and without incorporating application-specific processes. In this context, we also explore potential power savings through the elimination of redundant KPI transmissions for the first time, extending existing techniques for identical subscription removal and KPI selection. We achieve significant power consumption gains exceeding 87% of the overall RIC power consumption.

A testbed with open-source tools and commercial equipment is designed to investigate the timing of ML-based xApp operations in DDoS attack detection during the Random Access procedure. Results show high-precision and high-recall ML classifiers operating faster than the Medium Access Control (MAC) layer's Contention Resolution timer. For classifiers with over 97% precision and recall, XGBoost has the fastest execution time and the narrowest estimated run time probability density function.

Although the 5G New Radio standard empowers the mobile communication networks with diverse technologies such as Massive MIMO, mmWave deployments, and much more, some network functionalities still do not explore the potential of assembling Artificial Intelligence to their methodologies. The handover procedure is planned very similarly to in older 3GPP networks, based on simple power measurement comparisons and rudimentary parameter tuning, such as Time-To-Trigger and Hysteresis. This work develops and evaluates with simulations and real network data a new Deep Learning approach to support the handover triggering decision toward a data-driven procedure for next-generation networks. Our solution relies on predicting future samples of standard Reference Signals using Long Short-Term Memory Networks (LSTM) in the first stage. After, the predicted power samples are sent to a binary classification algorithm to identify if the time series will lead or not to a handover triggering. The results show a mean absolute error of around 0.6 dB predicting power signal samples and over 97% of accuracy, indicating the future handover trigger moment. Finally, we discuss possible use cases to implement our model, including Open RAN and MEC architectures.

In this work, we investigate the sum-rate performance of multicell and cell-free massive multi-input multi-output (MIMO) systems using linear precoding and multiuser scheduling algorithms. We consider a network-centric clustering approach to reduce the computational cost of the techniques applied to cell-free systems. We then develop a greedy algorithm that considers multiple candidates for the subset of users to be scheduled and that approaches the performance of the optimal exhaustive search. We assess the proposed and existing scheduling algorithms in both multicell and cell-free networks with the same coverage area. Numerical results illustrate the sum-rate performance of the proposed scheduling algorithm against existing approaches.