Performance comparison of LSTM-based handover prediction models for 6G MEC services

6G networks aim to offer users ultra-efficient access to networked services based on seamless integration of the network with low latency solutions such as the Multi-access Edge Computing. To achieve this goal, the 6G edge orchestration should support smart decisions based on the learned user mobility behaviour. In this paper, we focus on the handover prediction problem that is an essential component to implement proactive service migration in 6G networks, particularly when migration delays require reliable forecasts several seconds ahead. We analyse three novel models based on Long Short-Term Memory (LSTM) neural networks aiming to improve the handover prediction performances when working with a wider prediction horizon. The architectures include: a parallel multi-output model that produces all outputs simultaneously, a multi-output model with the addition of a Bahdanau attention mechanism that works by iterative prediction of future time points, and an ensemble of single-output models that incorporates previous predictions into future steps. Our results show that the single-output ensemble models provide stable results over the entire prediction horizon, although prediction time costs must be assumed. The iterative attention model achieves a balance in mid-range predictions, while the parallel multi-output model stands out for its excels in early-horizon prediction. These findings highlight a clear trade-off between predictive accuracy and computational cost, indicating that the MEC orchestrator should choose an architecture guided by the requirements of the targeted MEC service and the lead time available for proactive migration. • 6G MEC handover prediction with Bi-LSTM and extended horizon in dynamic urban areas. • Comparative analysis of handover prediction with multi and single output attention. • Balancing accuracy and resources between multi and single-output models. • Long-term handover prediction strategy recommending optimal modeling approaches based on prediction horizon length.