Energy-Efficient Computing Offloading Policy for MEC-assisted Power Sensor Network: A D3QN Approach

With the deep and close integration of massive renewable and controllable terminals into the next generation smart power grid, the amount of data and various applications with different quality requirements will increase greatly, which will be tough challenges for the power sensor network to provide real-time data collecting and processing, global monitoring, and quick response, etc. Meanwhile, due to the limitation in energy supplement and computing capacity, it is a second issue for the sensor network how to work in an energy-efficient way. In this paper, our work sets out to study a reinforcement learning (RL) based computing offloading policy in the mobile edge computing (MEC) assisted power sensor network. By jointly optimizing task scheduling and resource allocation, we propose a new scheme to optimize data latency and energy efficiency. The optimization problem is modeled as a Markov decision process (MDP), and Dueling Double Deep Q Network (D3QN) algorithm is employed to solve this MDP problem. Simulation results demonstrate the effectiveness of the proposed algorithm and its outperformance in reduction of data latency and energy consumption as well as convergence and stability, compared with other benchmarks.