Partial Discharge Pattern Recognition Of Transformer Based On Probabilistic Neural Network

Electric power is the lifeblood of the economic development in the modern society. In recent years, disastrous power cut accidents happened continuously all over the world. Most of these accidents are due to the power system fault.Transformer is the hub of power system. Its running state is straightly related to the dependability of the power system.International Conference on Large High Voltage Electric System(CIGRE) report indicates that insulation fault accounts for 51% of high voltage electrical fault. The main reason of insulation degradation is long-time of partial discharge(PD). It is necessary to apply partial discharge detection to transformer and evaluate the degree of insulation degradation, so that it can avoid insulation fault and ensure the reliable and safe operation of transformer.PD pattern recognition is the core of PD detection. It can recognize the type of insulation fault fast and help engineering technicians to determine the repair plan.Therefore, this dissertation about the transformer PD pattern recognition has larger theoretical value and practical meaning. The details are as follows.(1)This dissertation bases on ERA to build the test platform and adopts PCI-9814high-speed data acquisition card and LabVIEW software to build the PD signal acquisition system. Then this dissertation uses the signal acquisition system to acuisit the PD signals of four typical transformer partial discharge models. It uses the data which collected by a lot of experiments and then bases on PRPD to map the three-dimensional spectrum. Then the three-dimensional spectrum maps to two dimensional planes, after that, it generates gray-scale. Because moment feature can reflect the distribution of each pixel, it collects gray-scale moment feature as input features of pattern recognizer.(2)Partial discharge detection process has high real-time requirement and detected signal is easily interfered by random noise, which leads to many error samples. For the above limitations, the partial discharge pattern recognizer of transformer based on probabilistic neural network is proposed, which has a strongercapacity of additional sample, higher fault-tolerance and faster learning speed. In order to verify the performance of the proposed pattern recognizer, it is compared with back propagation neural network, extreme learning machine and naive bayesian recognizer. The simulation result reveals that, compared with other pattern recognizers, the proposed recognizer has higher accuracy.(3)The smooth factor of probabilistic neural network has important influence on recognition effect. At present, this factor mainly depends on empirical value, which requires complex calculation and experiment. To solve this problem, smooth factor can be optimized by using genetic algorithm, which has global search ability. The simulation result reveals that the recognition effect of propagation neural network has been greatly improved by optimization.