AutoMEC: LSTM-based User Mobility Prediction for Service Management in Distributed MEC Resources

The 5th generation of the cellular mobile communication system (5G) is in the meantime stepwise being deployed in mobile carriers' infrastructure. Various standardization tracks as well as research activity are investigating the exploitation of the very flexible 5G system architecture for customized deployments, meeting requirements of the vertical industry, such as for automotive, factory, or smart city. A very common base is a cloud-native development and decentralized deployment of the 5G system along with services in distributed resources per the Multi-Access Edge Computing (MEC) architecture to locate services topologically close to (mobile) users, e.g. along public roads, and to enable low-latency communication with local services. Automated management of such a distributed deployment in an agile environment is a prerequisite. This paper investigates the use of Recurrent Neural Networks (RNN) for accurate user mobility prediction in an automotive scenario. By the use of simulated vehicular traffic, a suitable RNN configuration using Long Short-Term Memory (LSTM) has been found, which provides accurate prediction results. Proof of value has been accomplished by an experimental decision algorithm, which balances the use of available distributed resources through service scale, migration or replication decisions while meeting mobile users' expectation on the experienced service quality.