Abstract

This thesis presents a unified framework for enabling sustainable cognitive networks through the integration of machine learning and energy-aware optimization. As networks evolve toward 5G and beyond, growing demands in performance, energy efficiency, and autonomous management call for intelligent, scalable solutions. This work addresses these challenges through a holistic approach spanning four layers: data generation, infrastructure optimization, distributed learning, and predictive control.

To enable AI-driven intelligence, techniques are developed at the data layer that leverage data-centric AI to generate high-quality synthetic 5G packet-level data and refactor real-world urban data into machine learning-ready, 5G-like flow-level traces. These methods mitigate data scarcity and heterogeneity, providing realistic and diverse data essential for robust network learning systems.

In the infrastructure layer, the placement of disaggregated 5G components, including Distributed Units, Centralized Units, and User Plane Functions, is formulated as a large-scale optimization problem. The proposed decomposition-based and heuristic approaches improve energy efficiency by up to 14\% while maintaining Quality of Service and responsiveness. Experiments in simulated 5G environments highlight the limitations of traditional peak-time-based planning.

For distributed learning, AFSL (Asynchronous Federated-Split Learning) and its energy-aware variant, AFSL+, are proposed to address client heterogeneity, achieving convergence time reductions of up to 13\%. These frameworks selectively engage participants, reducing energy consumption by up to 55\% without sacrificing accuracy and stability. To minimize communication overhead, Ada-AFSL is introduced, a dynamic compression technique that adapts to real-time bandwidth fluctuations. In realistic 5G and IoT scenarios, it achieves up to 82\% data reduction while preserving performance and enhancing generalization.

Lastly, ST-SplitGNN and ST-SplitGNN+ are developed as spatio-temporal split learning models for accurate traffic prediction and uncertainty-aware resource allocation. Learning process is partitioned between local and centralized components: at the edge, temporal encoders capture node-specific traffic patterns, while a centralized Graph Neural Network with a learnable adjacency matrix models time-dependent inter-node dependencies. These models enable proactive scaling policies that align reliability with sustainability goals.

Together, these contributions form a cohesive architecture for sustainable cognitive networks. By uniting optimization with machine learning, this thesis supports the development of intelligent, efficient, and environmentally responsible communication systems for the 5G era and beyond.