Machine Learning-based Direct Multi-Step Ahead Handover Prediction for Stand-Alone 5G Systems in Maritime Environments

Abstract:
5G introduces new capabilities that enable autonomous vehicles to be deployed in a variety of different scenarios. Maritime scenarios possess additional challenges compared to land-based scenarios, especially during the handover process. Handover prediction can be used to prepare for and possibly prevent these issues. However, in order to properly react to these prediction results, multi-step ahead prediction is necessary. Machine learning provides several tools to accomplish this task. In this work, a long short-term memory and eXtreme Gradient Boosting were trained on real measurement data from a research vessel. Both algorithms achieved a high prediction accuracy of about 90% when predicting just one second into the future, but when attempting to predict further into the future, XGBoost quickly drops off in its accuracy while the LSTM loses almost no accuracy.