Research On Power System Fault Diagnosis Based On Clustering And Recurrent Neural Network

With the rapid development of social economy and industrial production,the power system has become an indispensable part of the national infrastructure system.In recent years,power grid equipment has become smarter,and the scale of power grids has continued to expand.Power grid data presents the characteristics of large data volume,complex types of equipment used for recording,low frequency of failures,and unevenness.Traditional power grid fault diagnosis methods usually use system modeling,switch data analysis and other methods,and face new challenges in the rapid development of smart grids.Aiming at the phenomenon of incomplete faults,imbalanced categories and small samples,using data processing methods such as clustering and recurrent neural networks and deep learning methods,researches on power grid fault diagnosis are carried out.The main research contents are as follows:1)A fault diagnosis method based on semi-supervised clustering and recurrent neural network is proposed.Aiming at the time series data and unlabeled data existing in the power grid,firstly,the convolutional neural network is used to extract features for the samples,and semi-supervised clustering is performed on the sample feature sets to predict the label for the unlabeled samples.Then train the built recurrent neural network model to learn the context of time series signals,and finally use the trained model to diagnose the fault.Experiments show that this method can accurately assign labels to samples and improve the accuracy of diagnosis.2)A power grid fault diagnosis method based on GRU and oversampling is proposed.Aiming at the over-fitting phenomenon caused by the imbalance of label categories and the inconsistency of sample length,first interpolate the samples to obtain the same length,and then oversample the fault categories with a small number of samples to generate new samples to balance each fault category.The introduction of GRU network avoids the problems of gradient disappearance and long-term dependence,and the relatively simple internal structure can also ensure the efficiency of model training.The performance of the algorithm is verified by the simulation experiment on the grid fault data set.3)Propose an enhanced fault diagnosis method based on GAN.In order to generate samples under small sample conditions and extract sample features more accurately,this method first trains generative adversarial networks through real samples and generates sample augmentation data sets through the generator module in the network.After that,the encoder part of the stacked sparse autoencoder is used to encode the samples,extract the high-level features in the samples,and use the deep neural network to identify the fault features.Finally,the DC network fault data set is used to verify the effectiveness of the diagnosis method.