Prediction Method Research Of Transformer Fault Based On Extreme Learning Machine

The operation of the transformer condition is directly related to the safety of the power grid. Gases dissolved in transformer oil can reflect the operation of the transformer, but the oil dissolved gas analysis(DGA) method can only diagnosed the transformer fault in the moments of sample collection, and can’t predict transformer latent fault. Therefore, the research on forecasting method of gas dissolved in transformer oil is provided with important theoretical value and practical significance.First of all, this paper based on the analysis of the characteristics of transformer fault prediction method and its existing problems, the basic theory and algorithm of extreme learning machine was introduced in detail, and use it to transformer fault forecast, explore a new forecasting method of dissolved gas in transformer oil based on extreme learning machine.Then, a transformer fault prediction method based on regular extreme learning machine is given in this paper. This method is aimed at extreme learning machine in the case of limited samples will appear the phenomenon of excessive fitting, so we balance the empirical risk and structural risk by introducing the structure risk minimization principle and balance factor. This approach improved the generalization performance of extreme learning machine effectively under the premise of keeping fast training speed. Experiment shows that Regular Extreme Learning Machine has a better fitting and generalization performance.Finally, on the basis of the mixed kernel function, Kernel based Extreme Learning Machine and Particle Swarm Optimization(PSO) algorithm is studied, a new method of transformer fault prediction based on Particle Swarm Optimization Mixed Kernel Extreme Learning Machine is presented. The mixed kernel function is weighted by the Gaussian kernel function with local characteristics and polynomial kernel function with the global properties. In view of the kernel function parameter sensitivity problems, using Particle Swarm Optimization algorithm to optimize the parameters of extreme learning machine, gives the specific optimization process of parameters. Experiments show that transformer fault prediction based on Particle Swarm Optimization Mixed Kernel Extreme Learning Machine has better prediction effect.