Application Research On Transformer Fault Diagnosis Of Several Neural Network Methods

The power transformer is responsible for the power transmission and transformation in the power grid,which is the key point in power system,which will cause serious consequences in the event of failure.In order to find the potential faults of transformer ahead of time and remove failuresin the bud,we need to study the fault diagnosis of transformer.There would produce a lot of state data that closely relate with running state when the power transformers run,the dissolved gas in oil is the most commonly used state.The main types of transformer faults are thermal faults and electrical faults,a large amount of gas will be dissolved in the transformer oil when a fault occurs,its components and content are related to transformer fault type,transformer fault diagnosis method based on dissolved gas in oil is to find out the relationship and use this relationship to judge the fault type of the unknown data.Dissolved gas analysis in oil is an important means to monitor the safe operation of oil immersed transformer,this paper explains the relationship between dissolved gases and fault types through literature,aimed at the shortcoming that three-ratio method code is not complete and code limit is too absolute lead to applying to not all situations,studied the fault diagnosis method based on neural network.There are many different types of neural networks,we select the typical feed-forward BP neural network,extreme learning machine and deep learning as the main modelsin this paper,the simulation experiment shows that the performance of BP neural network is worst,extreme learning machine has the fastest learning speed but the model stability is the worst,deep learning can achieve the best learning performance but the learning time is long.Aiming at the limitation of random initialization weight of extreme learning machine,an improved model based on restricted Boltzmann machine is proposed in this paper,the network performance of the extreme learning machine is greatly improved by pre training the the weight matrix.