Mobility prediction via sequential learning for 5G mobile networks

Here, we present a mobility prediction framework for 5G mobile systems. Our work stems from the intuition that mobility in vehicular networks is highly correlated, and such correlation can be captured by advanced neural network designs to anticipate the users' point of attachment. To prove this, we combine Markov chains with recurrent and convolutional neural networks, training them on mobility trajectories estimated by the received radio signal from mobile millimeter-wave devices. The proposed framework is decentralized, i.e., user trajectories are independently learned by each base station. In this paper, various problems are pragmatically tackled and solved, such as dealing with imbalanced datasets, as some trajectories are under represented, and obtaining a mobility classifier whose accuracy increases as new mobility samples are collected.The proposed technique is assessed using emulated traces obtained through the SUMO mobility simulator for the city of Cologne. Numerical results show accuracies higher than 88% in the prediction of the next serving base station from 4 seconds before the handover is performed. Mobility (next base station) predictors like the ones presented here are key for network management purposes within 5G networks, e.g., to proactively allocate communication and edge computing resources.