Recognition And Classification Of Power Quality Disturbance Based On Deep Learning

The current rapid development of science and technology has accelerated the construction and development of power grids.The integration of distributed power generation and the access of nonlinear and diversified loads have brought new challenges to the stable operation of the power grid.In order to ensure the power quality of the power provided by the power grid,It is necessary to detect and classify the disturbances in the power system.This paper studies the classification of power quality disturbances to provide a basis for the management and protection of power quality.In order to overcome the shortcomings of redundancy and incomplete feature extraction of traditional power quality disturbances,a new method of power quality disturbance classification based on convolutional neural network is proposed.On the basis of the traditional convolutional neural network,the introduction of side output fusion structure,through the combination of convolution low-level,middle-level and high-level information for feature fusion,can better grasp the overall and local features of the signal,and effectively improve the classification accuracy.Use batch normalization structure and learning rate settings to optimize the neural network to avoid overfitting.Aiming at the problems of insufficient measured signals and uneven number of disturbance signals of different types,data enhancement processing is carried out.In view of the high network complexity of most current deep learning algorithms and the time-consuming and laborious identification and classification process,a new method of power quality disturbance classification based on ensemble learning and LSTM neural network is proposed.The Bagging algorithm integrates the classification results of multiple LSTM networks with differences,thereby effectively improving the generalization of the network.In view of the situation that there are less label data in the disturbance classification process and a large number of unlabeled power quality disturbance signals,the Bagging-LSTM network framework is used as the basic classifier,combined with the active learning uncertainty sampling strategy,and the selection is more conducive to improving the performance of the classifier The disturbance signal samples can achieve better classification performance with less manual labeling workload.In view of the lack of current power quality disturbance label data and the timeconsuming and laborious characteristics of manual labeling,the classification model is used to reduce the data demand of the classifier.Based on the transfer learning theory,the structure and weight parameters of VGG-16 trained on the Image Net data set are transferred to the problem of power quality disturbance identification.By migrating the parameters and weights of the convolutional layer and pooling layer of the model,the network optimization of the power quality disturbance signal depth feature extraction model under the limited number of training samples is completed.