Fault Identification Method For Inter-turn Short Circuit Of Power Transformer Based On Deep Learning

The development of industry has driven the growth of energy demand and improved the status of the power system in my country’s infrastructure system.As the core equipment of the power grid,the operation of the transformer directly or indirectly affects the security of the entire power system.Statistics show that among all kinds of transformer faults,the probability of inter-turn short-circuit faults in windings is the largest,accounting for about 60%-70% of the entire faults.At present,the inter-turn short-circuit diagnosis technology needs to be shut down and disassembled,which increases the operation and maintenance cost,and cannot find the potential fault of the transformer.The recognition of the characteristics of inter-turn short circuit is still based on the identification method of the circuit model,and the limitation of this kind of method lies in the need to establish an accurate equivalent model of the transformer product.Aiming at the identification of transformer winding inter-turn short circuit,this thesis proposes an intelligent diagnosis method for transformer winding inter-turn short circuit fault based on the deep learning classification framework.Firstly,the mechanism analysis of transformer winding inter-turn short circuit is carried out.The port voltage,current amplitude and transformer fault phase magnetic flux of transformers with different voltage levels are analyzed,and the applicable scope of the transformer winding inter-turn short circuit identification method based on deep learning technology is obtained.From the two aspects of time domain and frequency domain,the characteristics of voltage and current data of transformer ports are analyzed.On the basis of these characteristics,a method of making data sets is studied according to the requirements of data sets required for deep learning.Secondly,the transformer port voltage and current data set is constructed by the data set production method of the above research,and the convolutional neural network algorithm is used to perform automatic feature learning and pattern recognition on the data set samples.In the process of building the convolutional neural network,the structural characteristics and training process of the convolutional neural network are analyzed,and the input layer,convolution pooling layer,fully connected layer and output layer are designed according to the characteristics of the transformer port voltage and current data sets.The constructed convolutional neural network model is used to identify the inter-turn short circuit of the secondary winding of the transformer under the conditions of different voltage levels,different load rates,three-phase unbalance,and inter-turn short-circuit of the transformer winding at different positions.Finally,when the convolutional neural network classification structure identifies the transformer inter-turn short-circuit fault,the input signal characteristics of the internal parameters of the deep learning structure are the key factors to improve the accuracy.Taking the identification of short-circuits with different turns in the secondary winding of the transformer as an example,the identification effect was analyzed by changing the number and size of convolution kernels,the pooling strategy and the number of convolution pooling layers and other internal parameters of the convolutional neural network structure.The input signal characteristics such as the resolution and window size of the dataset image,the measurement accuracy of the transformer and the sampling frequency affect the identification effect.The research results of this paper supplement the basic theory of fault diagnosis of power transformers,and have broad application prospects in the field of intelligent fault diagnosis of power transformers.