Trajectory Prediction with Recurrent Neural Networks for Predictive Resource Allocation

Abstract

Trajectory prediction of mobile users plays a key role in making the plan for predictive radio resource allocation, i.e., determining which base stations alongside the trajectory of a user serve the user with how much resources. Predictive resource allocation in existing literature requires the prediction with second-level resolution and minute-level horizon. However, the trajectories predicted with existing methods are either too coarse-grained or with too short-horizon. In this paper, we strive to filling this gap by developing a recurrent neural network based trajectory prediction method. With proper network architecture and output structure, the proposed method can provide high-accuracy prediction with horizon of one minute. We investigate the performance of large-scale channel prediction with a perfect radio map. We also provide the statistics of the prediction errors for trajectory and large-scale channel gains, which is useful for the robust optimization of predictive resource allocation.