Research On Power Quality Disturbance Identification Method Based On Deep Convolutional Neural Network

With the exploitation of economy and society,lost of power electronic devices,nonlinear loads and distributed generation are extensively linked to the power system,which enlarges the nonlinearity of the power system,and the disturbances of power quality are more sophisticated and changeable,which severely threatens the normal operation of the power system.Therefore,finding a method that can quickly and accurately identify various disturbances of power quality has evolved into a research focus,which is important to the safe and stable operation of the power system,the normal and stable work of power load,and the orderly exploitation of economy and society.From the angle of convolutional neural network,this paper studies the recognition method of power quality disturbance,the main contents are as follows:(1)This paper describes the research background and meaning of power quality disturbance identification,the feature fetch and classification methods used in traditional power quality disturbance recognition approachs,and the research present situation of convolutional neural networks at domestic and foreign.The relevant domestic and foreign standards of power quality are elaborated,the causes and hazards of various disturbances are briefly described,and one standard signal,seven single disturbances and six composite disturbances are modeled according to the standards,and the data sets required for this paper are structured.(2)In view of the traditional convolutional neural network that tends to degrade or even not converge when the model depth increases,the power quality disturbance recognition method rely upon deep residual network is researched,and the residual block is used for adding the network depth.The structure and parameters of the network are determined by horizontal comparison experiments,and simulation experiments indicate that compared with traditional CNN,Res Net can practically improve the network degradation problem,convergence speed is faster,classification accuracy is higher,and noise robustness is better.(3)For the difficulties that the traditional convolutional neural network has a single convolution kernel,can only withdraw fixed-scale features,has inadequate feature extraction ability,at the same time,the feature map weight of each channel dimension is the same,which cannot prominent the effective features,multi-scale convolution and efficient channel attention mechanism are combined to improve the feature mining ability of the model,and considering the shortcomings of the efficient channel attention mechanism that easily loses feature information in the process of downsampling,in the feature extrusion process of efficient channel attention mechanism,the hybrid pooling composed of global maximum pooling and global average pooling is used to substitute the original global average pooling,and a power quality disturbance recognition method rely upon channel selection multi-scale fusion residual network(1DIRMECA-CNN)is constructed based on the residual idea of residual thinking.Through simulation experiments and actual data verification,1DIRMECA-CNN has stronger deep feature extraction ability than deep residual network and deep convolutional neural network,and power quality disturbance signals can be classified more precisely.(4)For the difficulties that there is lots of noise interference in the actual power system,which makes the accuracy of power quality disturbance signal classification low,a one-dimensional deep residual shrinkage network is introduced,and on the basis of the traditional residual network,attention and soft threshold mechanisms are added,and the two cooperate with each other to set the threshold adaptively to realize the soft thresholding of each feature channel and remove redundant features such as noise;On this basis,dynamic convolution is introduced,multiple parallel convolution kernels are dynamically aggregated according to input attention,the convolution kernels are adjusted adaptively,at the same time,the feature distill capabilities of the model is raised,and a dynamic convolution depth residual shrinkage network(1DYDRSN)is formed to categorized the power quality disturbance signal with strong noise.Simulation experiments and actual data verification show that 1DYDRSN can still classify power quality disturbance signals well in strong noise environment,which provides a new idea for detecting power quality disturbance in strong noise environment.