V2V Task Offloading Algorithm with LSTM-based Spatiotemporal Trajectory Prediction Model in SVCNs

Abstract

In this paper, in order to fully investigate the great potential of increasingly powerful vehicles and their predictable mobility in improving dynamic network performance, we develop a V2V task offloading algorithm for SVCNs (Software-defined Vehicular Cooperative Networks). First of all, a three-layer SVCNs architecture is built with vehicle edge layer, SDN control layer and cloud layer from bottom to top, moreover, at SDN control layer, there are multiple SDN controllers responsible for data collection and resource scheduling, and a dual LSTM-based vehicle trajectory prediction model is also designed with the aim of achieving a spatiotemporal trajectory prediction process. We then formulate an optimization problem and further show our V2V task offloading algorithm design, which consists of a distributed V2V (dV2V) offloading sub-algorithm and a centralized V2V (cV2V) offloading sub-algorithm. Finally, we test our trajectory prediction model with TensorFlow, and simulate our V2V task offloading algorithm on NS3 platform, results show that not only our prediction model obtains the highest accuracy, but also our V2V task offloading design outperforms the comparison schemes in terms of average service time, task success ratio and algorithm convergence with various vehicle density and motion modes.