With the development of 5G communications and Mobile Edge Computing (MEC), a variety of new demands for power Internet of Things (PIoT) have emerged. On the one hand, these new PIoT applications usually request for high quality of service (QoS) guarantee. On the other hand, service providers are desired to have elastic framework for diverse service level agreements (SLA). The feature technologies of network slicing and MEC have constituted a practical viable framework for solving these challenges. This article first defined a metric to measure the reliability of PIoT services. After that, a deep reinforcement learning based network slicing approach is proposed to jointly optimize the computing and communication resources. The proposed approach ensures the latency and reliability of the PIoT services while minimizing their energy consumption. Through simulation experiments, the slice management approach is demonstrated to outperform the traditional baseline approach.