Recognition Of Transformer Discharge Phase Pattern Based On Generative Adversarial And Convolutional Networks

The safe operation of transformer is the key to ensure the reliable power supply of power system.It is a common method to judge partial discharge type of transformer by discharge distribution phase(PRPD)graph in power enterprises,but it requires maintenance personnel with high experience.Deep learning algorithms shine in the field of image recognition,but PRPD atlas data sets often have problems such as data imbalance and small amount of data,which cannot meet the research premise of deep learning.Therefore,this paper studies PRPD map recognition based on generative adversarial network and convolutional neural network.(1)This paper is designed an extension method of PRPD spectrum data set based on residual-generative adversarial network.Based on the GAN generation model,this method is added input and mapping between each two deconvolution layers in a "jump link" mode to enhance model stability,improve the problem of mode collapse,alleviate gradient disappearance effectively,and improve image quality and diversity.Used spectral normalization instead of batch standardization,the discriminator can satisfy Lipschitz continuity and maintain the stability and efficiency of training effectively.The dual time scale update rule is ensured that the generator will take smaller steps to learn,making the training more stable.(2)This paper is constructed a lightweight convolutional neural network algorithm and applied it to PRPD map recognition after data expansion.The algorithm is introduced deep separable convolution to reduce the number of network parameters,improve the computational efficiency,and make the model lightweight.Swish function was used to replace Re LU in the activation function to eliminate the influence of neuron death,and adjustable parameters were used to realize dynamic learning of the activation function.The strategy of "convolutional layer + pooling layer" is adopted to build CNN network structure,which tries to meet the recognition requirements of maps of different sizes.The convolutional neural network is trained and tested using partial discharge sample data set.Experimental results show that the average recognition accuracy of the improved convolutional neural network is 97%.