Optimisation foundations for AI-based Hypercell operation

In collaboration with Huawei, the IP led the development of the optimization foundations for AI-based Hypercell operation, addressing how integrated terrestrial/non-terrestrial networks can be configured to maximize energy savings while preserving user quality of service. As an initial contribution, the project formulated a theoretical optimization framework in which a wide-area non-terrestrial layer offloads traffic and enables the shutdown of terrestrial base stations during low-load periods [E2.1]. Although this work considered LEO satellites, the same optimization principles apply to Hypercell architectures based on HAPS or drone-assisted coverage layers. The framework jointly optimizes user association, bandwidth allocation, transmit power control and base-station activation, leading to the BLASTER algorithm. Simulations showed terrestrial energy-consumption reductions of up to 67% during low-traffic periods, while preserving or improving network performance, providing a basis for the project and subsequent work.

Building on this basis, the IP initiated the development of AI-based optimization methods for the Hypercell digital twin developed in SO3 [E1.10]. Bayesian Optimization was studied as a sample-efficient method to explore energy-efficient configurations in expensive network simulations, while reinforcement-learning approaches, including PPO and Actor-Critic methods, were investigated for sequential shutdown and activation decisions under dynamic traffic conditions. This work also analysed the performance requirements of exchanged AI models, such as traffic and SINR predictors, relating model accuracy to network-level KPIs, update frequency and signalling overhead.

This activity was strengthened through collaboration with Universitat Pompeu Fabra and Huawei, where High-Dimensional Bayesian Optimization, Reinforcement Learning and Transfer Learning were developed and validated in related cellular-network optimization problems, including UAV corridor design, mobility management and antenna configuration [E1.2, E1.3, E1.4, E2.2, E2.3, E2.4]. These studies provided methodological foundations for SO4 by demonstrating how BO/RL-based techniques can optimize high-dimensional, site-specific and data-driven cellular scenarios, and how transfer learning can accelerate convergence across related network configurations.

Overall, SO4 reached an estimated 85% completion. The project delivered the theoretical optimization formulation, first algorithmic solutions and methodological basis required for AI-based Hypercell optimization. The full end-to-end integration of BO and RL into the Hypercell digital twin remains ongoing beyond the project, while the first machine-learning-based Hypercell optimization results are being consolidated in a dedicated publication currently under preparation [E1.10].