Implementation of deep learning based power system diagnosis in edge computer

Abstract

The popularity of the power grid has been around for decades, the insulation in the grid has been gradually aging over time. The broken of insulation layer will increase the risk of its breakdown, which may have a huge impact on the entire power system, resulting in an inestimable economic loss. Partial discharge phenomenon is the precursor of power system failure. On the one hand, this phenomenon is the symbol of insulation deterioration. On the other hand, it will destroy the insulation structure and degrade its performance. Therefore, detection and identification of partial discharge signal is an important method to prevent power system faults. Nowadays, as the development of machine learning, more and more deep learning algorithms are applied in PD detection, and most of them have very good performance. However, it is not suitable to send the data to the cloud server for calculation because the data of PD signal is very large and involves national and enterprise secrets. Edge computing is a solu tion to this problem. By offloading some computations to the edge, the energy consumption required for communication can be greatly reduced, the response can be obtained faster, and the privacy of the data can be protected very well. Therefore, using deep learning algorithm to detect PD signal on edge computer is becoming a potential development direction. In this dissertation, a real-time PD detection system based on edge computer is presented. The data acquisi tion, classification and identification of PD are realized by pipeline method, and the results are uploaded to the back end in DE-10 SoC FPGA edge computer board.