**1. Introduction**

In recent years, along with the development of the Internet of Things (IoT) technology, it has become easier to connect mobile devices to the Internet [1–3]. In particular, IoT-based sensor devices, which have low computing performance, can overcome their computational limitations with the help of cloud systems [4,5]. However, the explosive growth of IoT data has resulted in increased traffic load in networks and cloud systems. This overload reduces the quality of experience (QoE) of the services and can result in network blackout, which shuts down the network system [6,7]. To solve these problems, the mobile edge computing (MEC) technology, which is a type of radio access network (RAN) with cloud computing capabilities, has been developed to assist resource-constrained IoT

devices [6,7]. In an MEC environment, the MEC server computes the offloaded workload from the IoT devices and charges the bill accordingly. This paper proposes an energy harvesting (EH) MEC system, which enhances the survivability of the MEC through energy harvesting if the MEC is installed in a remote area where grid power supply is not available. In the EH MEC system, the MEC server is powered by renewable energy resources (RER) such as photovoltaic or wind turbine resources. Therefore, it is possible to establish a system that deploys IoT sensor nodes and collects information in places where it is difficult to install electricity facilities such as deserts or unmanned islands. However, the EH MEC system has important challenges in terms of stability. On the one hand, since the energy supply from RERs is uncertain with respect to weather or time, the EH MEC system cannot achieve stable energy unlike conventional grid powered MEC. On the other hand, if EH MEC operation only considers maximization of the system performance, system black out will occur since the power supply of EH MEC is unstable, i.e., EH MEC has to consider battery stability [8]. Thus, EH MEC reduces the energy consumption when the amount of harvested energy is not sufficient even if the system performance is decreased.

This paper proposes an EH MEC scheduling algorithm that considers the battery stability. Figure 1 shows the proposed MEC system. As shown in the figure, the IoT devices transmit the offloading requests to the MEC system. Then, the MEC system determines the admission of requests (referred to as offloading scheduling). If the offloading request is accepted, it is executed by the MEC server. Otherwise, it is transferred to the cloud system. In addition, the MEC server determines its computation frequency based on the battery state-of-charge (SoC) (referred to as MEC scheduling). According to circuit theories, the CPU power is dominated by the dynamic power, which originates from the toggling activities of the logic gates inside the CPU [9]. Thus, in this paper, we assume that the power consumption of the computation offloading can be handled by CPU frequency scheduling such as dynamic voltage frequency scaling (DVFS) [9,10]. In the proposed system, if the SoC of the battery is sufficient, the MEC server raises the computation frequency for faster offloading service. Otherwise, the MEC server will lower the frequency to ensure system stability, i.e. to avoid blackout.

**Figure 1.** Architectural view of the proposed EH MEC system, consisting of RERs, energy storage system (ESS), a base station, MEC, etc. The IoT devices send their computational workloads to the base station. If the computation can be performed by the EH MEC system, it is offloaded to the MEC. Otherwise, it is offloaded to the cloud system.

The contributions of this paper are summarized as follows:


The rest of this paper is organized as follows. In Section 2, we review related works on EH MEC systems. In Section 3, the scheduling algorithms are proposed, and the corresponding optimization problem is presented. The design of the greedy algorithm, which can obtain approximate optimal solutions for offloading scheduling, and a low complexity MEC frequency scheduling algorithm based on Lyapunov optimization are provided in Section 4. Section 5 presents the performance evaluation. The conclusions of this paper are drawn in Section 6.
