Research On Power Quality Detection Method Based On Deep Learning

With the large-scale application of power electronics technology,distributed generation and the wide access of many precision loads,power grid power quality is facing more and more new challenges.Complex and volatile power quality disturbances continue to threaten the safety of power users.Power quality problems in power grid have become the focus of research.In the study of power quality problems,the correct identification of power quality disturbance signals helps to reproduce and trace the disturbance,thus realizing the analysis and governance of disturbance,which is of great significance to ensure the safe and stable operation of the power system and improve the power supply quality of the system.From the perspective of deep learning,this thesis has carried out research on the classification and recognition of power quality disturbances based on deep learning algorithms.The main work of the thesis is as follows:The relevant standards of power quality are expounded,and the single and compound disturbance signal models of power quality are built.The causes and hazards of seven kinds of single power quality disturbance signals and five kinds of composite disturbance signals are introduced,and the 12 kinds of disturbance signals are simulated to obtain the disturbance signal samples,which provides the feature data set for the identification and classification of power quality disturbances by various deep learning algorithms.The power quality disturbance identification scheme based on deep residual network is studied.The residual module is used to deepen the structure of convolutional neural network,so as to solve the degradation problem of deep classification network.The influence of parameter configuration of deep residual network model on disturbance classification and identification is explored,and the deep residual network structure suitable for power quality disturbance identification is determined.The results show that the deep residual network can improve the network degradation problem of the existing network,and has higher disturbance recognition accuracy and noise robustness,and the convergence speed of the network is faster.The power quality disturbance identification method based on deep separable convolution neural network is studied.Separable convolution is used to reduce the computation of network model while ensuring the accuracy of disturbance signal recognition.The deep separable convolutional neural network and standard convolutional neural network are used to classify and identify 12 kinds of disturbances.The parameters,calculation amount and accuracy of the two network models are compared.The results show that the parameters and calculation amount of the deep separable convolutional neural network are lower than those of the standard convolutional network,and have higher noise robustness and accuracy.The data enhancement scheme based on W-GAN(Wasserstein distance generation confrontation network)is studied.Using the adversarial learning ability of W-GAN,the problem that the generalization performance of recognition network decreases due to the insufficient number of original disturbance signals is solved.The generator and the discriminator conduct multiple confrontation training,and gradually learn the real probability distribution of power quality disturbance signal.The trained generator is used to enhance the small sample data of the disturbance signal,and the real disturbance data and the generated disturbance data are combined as training sets to test the network.The example results show that W-GAN can effectively learn the probability distribution of the disturbance samples and improve the generalization ability of the model.In summary,this thesis studies two kinds of deep learning methods for power quality disturbance identification.Starting from solving the degradation problem of deep network and reducing the calculation amount and parameter number of network model,this thesis proposes deep residual network and deep separable convolution network,which have high identification accuracy.In view of the possible insufficient disturbance samples,a data expansion scheme based on W-GAN is proposed,which provides a new research idea for the identification of small sample disturbance data.