Research On Transformer Fault Diagnosis Based On Extreme Learning Machine And Its Deep Structure

At present,most of UHV/UHV transformers in China are equipped with oil chromatographic on-line monitoring devices.Based on oil chromatographic monitoring data,fault diagnosis of transformers has practical significance and is very important for the safe and reliable operation of power systems.On the basis of exploring the extreme learning machine and its depth structure,this paper attempts to apply the improved algorithm of extreme learning machine to transformer fault diagnosis based on DGA,in order to further improve the timeliness and accuracy of transformer fault diagnosis and better maintain transformer operation.Specifically,the following research has been done:(1)A transformer fault diagnosis algorithm based on NSGA2 optimized regular extreme learning machine(NSGA2-RELM)is proposed.The algorithm first uses NSGA2 to optimize the regular extreme learning machine,transforms the parameter optimization problem of the regular extreme learning machine into a multi-objective optimization problem,and then initializes the regular extreme learning machine with the optimized parameters to obtain a more stable network model.Finally,it is applied to transformer fault diagnosis.The algorithm aims at reducing the experience risk and structure risk of the original extreme learning machine.(2)A transformer fault diagnosis algorithm based on deep denoising extreme learning machine(DDELM)is proposed.Firstly,ELM and denoising autoencoder are combined to construct the basic unit-denoising autocoding extreme learning machine;secondly,the basic unit is stacked to construct feature extractor-deep denoising extreme learning machine;finally,the conventional ELM is added to the back of the feature extractor as a feature classifier to classify the extracted features.This algorithm combines ELM,denoising autoencoder and deep learning to extract features quickly and accurately from noisy DGA data.(3)Using the actual data of transformer faults in power plants to test two algorithms above to verify the effectiveness of the two algorithms in transformer fault diagnosis.It is proved that the two improved algorithms of extreme learning machine proposed in this paper have a certain application prospect in fast and efficient transformer fault diagnosis.