Towards smarter cellular networks management : from handover forecasting to robotic aerial base stations deployment

Abstract

Mobile cellular networks serve as the backbone of modern connectivity, supporting billions of users and facilitating communication and data exchange worldwide. As 5G subscriptions continue to grow and data-intensive applications proliferate, mobile operators face the dual challenge of managing the exponential increase in data traffic while guaranteeing a high quality of service (QoS) and quality of experience (QoE). The scale of these networks, combined with the need for real-time decision-making, renders traditional rule based and static management approaches fundamentally inadequate. Network operators are increasingly confronted with an environment where millions of parameters interact simultaneously across thousands of base stations, making manual or heuristic based management both impractical and suboptimal. This growing complexity has catalyzed a paradigm shift toward Artificial Intelligence (AI) and data-driven techniques as the cornerstone of next-generation network management. Among others, two aspects are to be considered in this topic, namely mobility management and traffic load management. Regarding the former, handover (HO) management which can encompass handover minimization and handover forecasting is critical. HO procedure is the process of transferring an ongoing communication session from one base station to another, which is pivotal for HO management. Each handover event introduces latency, increases signaling overhead, and carries a risk of connection failure, and the frequency of handovers is amplified dramatically by network densification, millimeter-wave communications, and heterogeneous deployments. Traditional handover mechanisms rely on fixed signal-strength thresholds triggering handovers once degradation is already detected. In dense and heterogeneous networks, this reactive behavior can lead to excessive unnecessary handovers, handover failures, and ping-pong effects that degrade throughput and increase signaling overhead. Regarding the latter, traffic load management encompasses all mechanisms aimed at monitoring, controlling, and optimizing the distribution of traffic across the network infrastructure. It can include load balancing, capacity planning, and traffic hotspot management among others. Traffic hotspots management specifically addresses the spatial and temporal concentration of demand in localized areas. This aspect is critical in maintaining consistent QoS across the network, particularly in dense urban environments where traffic is inherently heterogeneous and bursty. This thesis investigates AI-driven cellular network management as a paradigm to enhance network performance to address the shortcomings of the traditional methods, focusing on three dimensions. In our first contribution, we design a HO forecasting framework, which is a key component in HO management and consequently, mobility management. Particularly, we propose a framework relying on a combination of Graph neural Network (GNN) and Long short-term Memory (LSTM), named HGC-LSTM to forecast the number of handover occurrences among base stations over time. This framework captures the spatial-temporal dependencies in the network to improve the prediction. Said prediction is used to propose a dynamic resource allocation scheme in a virtualized Radio Access Network (RAN), thus constituting an anticipatory approach. Considering that HGC LSTM outperforms the chosen baselines, our findings when it comes to the handover resource allocation reveal that it is more advantageous in comparison to a traditional (static) approach. As for our second objective, we focus on defining, discovering and predicting mobility hotspots which highlight user mobility patterns. More specifically, we first propose a definition for what a mobility hospot is, due to a lack of an accurate definition in the literature. Then, we leverage the predicted hotspots for Robotic Aerial Base Stations (RABSs) deployment to showcase the relevance of mobility hotspots prediction in cellular networks. This aspect of mobility management serves in this case, to optimize RABSs placement in order to minimize their respective traveled distance which can be extended to ensuring energy efficiency. As for our final objective, we delve into mobility in a military tactical network, where the needs in terms of seamless connectivity are crucial. Considering our previous objectives, we deploy RABSs to minimize the number of handovers in a network including mobile groups of users and individual users. Our contribution introduces GCN-DDQN, a framework combining a heterogeneous Graph Convolutional Network (GCN) encoder with a Double Deep Q-Network (DDQN) for cooperative multi-RABS repositioning. The GCN encoder models the cellular network state as a heterogeneous graph that is fed into a DDQN that selects discrete repositioning actions for the RABSs, guided by a composite reward function combining a weighted handover penalty with a proximity bonus. We unveiled that a graph-based state representation is essential for the agent to learn effective RABSs positioning policies and consequently contribute to the minimization of the number of HOs.

Date	21 Sept 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Diala Naboulsi (Supervisor) & Razvan Stanica (Co-supervisor)