A Deep Learning Framework for Joint Handover Prediction and Path Stability Estimation in 5G/6G Networks

الملخص

The high mobility density of 5G and 6G networks, including vehicular systems, requires smart handover and imposes unparalleled demands on the network mobility management. The traditional handover mechanisms, which are largely reactive and inherently threshold-based, are inefficient to operate in dynamic environments and lead to severe impact, resulting in ping-pong effects, radio link failures, and inefficient use of resources. The paper presents the deep handover mechanism, a new end-to-end spatio-temporal deep learning model, which aims to orient the joint optimization of handover prediction and the following estimation of path stability. Our architecture represents a novel combination of a bidirectional long short-term memory (Bi-LSTM) network to learn complex temporal interactions in user mobility, with a Graph Attention Network (GAT) to learn the spatial interaction between the cellular topology. These multi-modal features are dynamically weighed by a special mechanism of attention-based fusion. Decades of testing on a large-scale dataset of 5.2 million network samples show that our proposed work obtains a state-of-the-art level of handover prediction accuracy of 92.4 %, even stronger than popular models such as LSTM-GRU hybrids (91.2%) and transformer-based models (90.8%). Moreover, it approximates stability of the paths with an average error of 1.3 seconds, and the inference computation time is only 12.8ms, which proves its applicability in real time. This publication introduces a new proactive paradigm of intelligent mobility management in heterogeneous networks.