Research On Fault Diagnosis Method Of AC/DC Power System Based On Deep Learning

The distribution of power generation energy and load in our country is extremely unbalanced,and long-distance transmission projects exist widely in our country.Compared with traditional AC transmission,DC transmission has incomparable advantages in long-distance transmission.Therefore,UHVDC transmission technology has been greatly developed in recent years,which has promoted the formation of AC-DC hybrid grid pattern in our country.An efficient power system fault diagnosis scheme is the premise for safe and stable operation of the power grid.The power system fault diagnosis scheme based on the switching value is very likely to misjudge the fault components when the protection and circuit breaker misoperate,the leakage and misalarm about alarm information occurs and so on.Electrical information has greater advantages than switching information in accuracy,completeness and fault tolerance,and with the development of power system monitoring system,a large amount of fault electrical quantity information can be obtained at the dispatching end.Therefore,this paper takes electrical quantity information as fault feature set and carries out research on fault diagnosis scheme of AC/DC hybrid power system based on deep learning algorithm.The main work of this paper is as follows:The research background and significance of fault diagnosis for AC/DC hybrid power system are investigated.The existing fault diagnosis methods of power grid are introduced.The current research status and characteristics of deep learning algorithm in power grid fault diagnosis are analyzed.In order to obtain a large number of fault samples,a programmable fault module is built on MATLAB/Simulink platform.Through interactive simulation of editor and Simulink,batch simulation of faults is realized.A total of 14,418 fault samples are obtained,which provides a good feature data set for deep learning algorithm.A fault diagnosis method based on SSAE(Stacked Sparse Autoencoder)and DBN(Deep Belief Network)classification model is studied.SSAE is used to reduce the dimension of the original fault sample data,which reduces the complexity of the data and facilitates DBN to classify the data.In this paper,SSAE dimension reduction effect is compared with SAE(Single Hidden Layer Sparse Autoencoder)and PCA(Principal Component Analysis)respectively.Through comparison,it is found that SSAE is better than SAE and PCA in reducing dimension of data.Finally,SSAE-DBN method is applied to fault diagnosis.The results show that the fault diagnosis accuracy rate of the proposed method is higher than that of traditional machine learning algorithm,and it is almost unaffected by fault type,fault location and fault transition resistance.The fault diagnosis method of AC/DC hybrid power system based on convolution neural network is studied.Convolution neural network has strong feature extraction ability,which can extract features from different aspects of data,so it can efficiently extract and classify the complex fault data in this paper.The influence of convolution neural network parameter configuration on fault diagnosis accuracy is explored.Through analyzing the influence of different parameter configurations on the accuracy and conducting many experiments,the convolution neural network structure suitable for fault diagnosis is finally determined.The convolution neural network is applied to the fault diagnosis of AC/DC hybrid power system.The results show that the convolution neural network has very high fault diagnosis accuracy,and the diagnosis accuracy for fault types and fault lines reach 99.93% and 99.95% respectively.Using t-SNE algorithm to reduce the dimension and visualize the convolution layer output,it is found that convolution neural network has the feature of extracting features layer by layer.Aiming at the problem that the generalization ability of discriminant model decreases due to the small number of fault samples,the DCGAN(Deep Convolution Generative Adversarial Networks)is used to expand the capacity of data samples.The convolution neural network is introduced into the Generative Adversarial Networks to improve the stability of the network.Through the adversarial study between the generator and the discriminator,the generator learned the probability distribution of real fault samples.The trained generator is used to generate new fault samples,and then the convolution neural network is used as a discriminator and classifier to calculate the correctness of generated data and the accuracy of diagnosis of test set data.Experiments show that the DCGAN model effectively learns the probability distribution of fault samples and improves the accuracy rate of fault diagnosis.