The Research Of Magnetizing Inrush Identification Method In Transformer Based On Wavelet And Neural Network

With the development of domestic power industry, new requirements of large transformer protection are presented. In the course of settling new types of transformer microcomputer protection, it is still a key and difficult problem how to prevent protection miss-operating caused by magnetizing inrush when transformer suddenly is closed in no load. The reliability of protection would be enhanced if magnetizing inrush is discriminated. So, this thesis analyses and compares some discriminating methods applied in practice or described in related literatures at present. On the basis of considering comprehensive factors, a method to discriminate between internal faults and magnetizing inrush based on wavelet analysis and dynamic architecture neural network is proposed.As the wavelet transform possesses the properties of multi-scale analysis and the time-frequency local characteristic, it is particularly well adapted to process sharply changing signals and to extract characteristic of these signals. When a smoothing function's first derivative is adopted as wavelet function, the local maxima of wavelet transform modulus detect the location of signal's singularities, In this thesis, these characters are not only applied to measure the magnitude of dead angles of the dissymmetry inrush but also the symmetry inrush in three-phase transformer. Measure precision is improved to some degree.The NN discriminating methods adopted in this thesis is different from the former based on NN. Firstly, second harmonic component ratio and dead angles of two phase inrush's dispersion in three-phase transformes are acted as input variable. Secondly,the method applies improved algorithm based on the original algorithm of multi-layer forward back propagation network, that is to say, adding last variational effect of weight value and bias value to this time and making use of variable learning rate. At the same time, this method also adopts dynamic form in the number of hidden floor node. Based on the above two solutions, the structure of NN is optimized by samples training. The simulative test results show that the training speed of NN model designed by this thesis is faster and theidentification of magnetizing inrush based on this method is effective.