Study On Recognition Of Complex Power Quality Disturbances Based On Deep Learning

With the wide application of new energy power generation and the increase in types of non-linear power loads,various power quality disturbances are increasing.Disturbances identification is the foundation of monitoring,analysis and governance of power quality,and has important practical significance for ensuring the safe,efficient and stable operation of the power grid.This paper systematically studies the complex disturbances recognition method based on multi-task deep learning classification,and conducts in-depth research on the identification characteristics,incomplete S-transform feature extraction method,multi-task disturbances classification modeling,and deep learning classifier design.The details are as follows:(1)The simulation models of single disturbances and composite disturbances are established,and the single disturbances are divided into four categories according to their characteristics.The basic principle of S-transform is studied and the identification characteristics of various disturbances are visually analyzed.An incomplete S-transformation method is proposed to transform only at several important samples,which can significantly reduce the time complexity of S-transformation while retaining important identification information,and the two triple disturbances are used as examples to verify the method.(2)Multi-task classification,a new complex disturbance recognition modeling method,is proposed for the first time.A multi-task label coding scheme and multi-task classification loss of complex disturbances are specifically implemented.The comparison results with multi-label classification and multi-classification verify multi-task classification can better model the label correlation of comlex disturbances,and the recognition effect is better.(3)An incomplete S-transform feature extraction technology is proposed to extract disturbance features.The basic principles of deep feedforward network(DFN)are studied,and a specific structure is proposed with dropout regularization to prevent overfitting.The experiments verify that the method can accurately identify various disturbances,which is more accurate than the traditional methods.(4)The basic components of convolutional neural network(CNN)and techniques such as batch normalization,dilated convolution,and residual blocks are studied to design more effective structures.A multi-dilation coefficient composite residual block is proposed,and a one-dimension CNN classifier structure is designed,which directly takes the signal as the input and complete the recognition end-to-end.The experiments indicate that the method has higher accuracy and good application prospects,compared with DFN and some existing methods.