Task Prediction Based Computation Offloading over Multi-UAV MEC Network

In mobile edge computing (MEC), unmanned aerial vehicle (UAV) acts as a base station that can quickly provide communication and computing services for areas with a limited communication infrastructure. However, most of offloading studies neglect the deep learning algorithm to understand the dynamic changes of task in different time slots. Moreover, in large-scale computing offloading, the UAVs deployment only considers position optimization, not number optimization. Considering the limited energy of terminal device (TD) and UAV, this paper proposes a joint Task Prediction (TP) and Differential Evolution (DE) optimization framework, called TpDeRas, to reduce the total energy consumption over multi-UAV MEC system. To predict the future task set, we first use a TP algorithm based on distributed long short-term memory (LSTM) to achieve task prediction for different TD. Based on TP results, the optimization problem is divided into UAVs deployment subproblem and resource allocation subproblem. UAVs deployment optimization needs to consider number and position. We propose an adaptive DE algorithm to optimize the UAVs deployment. Based on TP results and UAVs deployment, we uses efficient greedy algorithm to optimize resource allocation. Experimental results show that TpDeRas approach greatly improves performance compared to other algorithms.