LSTM based Proactive Access Point Selection and Mobility Load Balancing for Ultra-Dense Networks

To cater for the growing demand of capacity, Access Point (AP) densification is a promising solution. Within an Ultra Dense Network (UDN), a mobile User Equipment (UE) can associate with one of many candidate access points. Association with an optimal AP, from multiple eligible APs in proximity of a UE, while avoiding unbalanced UE distribution, becomes a challenge. Legacy Mobility Load Balancing (MLB) methods are usually reactive by design which hinders their efficiency to solve time restrained MLB problem. In this paper, an improved MLB framework is proposed, which utilises Recurrent Neural Network with Long Short Term Memory (RNN-LSTM), for users' temporal and spatial mobility prediction. It predicts AP's future load based on predicted distribution of UEs and finally optimises network's load in advance via a load aware AP to UE association based Load Balancing (LB) technique. This MLB framework optimises load pre-emptively by visualising congestion beforehand. The simulation results suggest that the performance of our MLB framework is superior to other existing algorithms that address time-constrained LB problem incurred by high mobility.