Study On Short Circuit Fault Type Identification Of Distribution Network With IIDG Based On Intelligent Algorithm

With the promotion of renewable energy in the power system and the development of smart grid,the huge amount of electrical information in the distribution network with DG needs to be mined and utilized urgently.The rapid development of artificial intelligence algorithm makes it possible to realize the fault type identification of distribution network with DG based on massive data,which is conducive to the safe,stable and reliable operation of distribution network and the construction of the new power system based on renewable energy.However,there is a lack of research on fault feature extraction of distribution network with DG and intelligent identification algorithm under the condition of unbalanced samples.In view of the above shortcomings,this paper studies the intelligent algorithm of distribution network fault type identification with DG from two aspects of feature extraction and intelligent algorithm.Aiming at the problem of feature extraction of distribution network with DG,the transient and steady-state features of distribution network with DG are obtained by basic theory analysis.Combined with the mathematical model of the output characteristics,DG is equivalent to VCCS to analysis the sequence network low resistance fault.Considering its four different output characteristics,the steady-state short-circuit current is calculated by equal modulus method elimination,and the steady-state characteristics of low resistance fault are obtained by the quantitative comparison analysis under three conditions of DG or no DG,different DG permeability and different output characteristics.Based on the transient process of high resistance fault,wavelet packet decomposition is used to extract the wavelet packet energy ratio of zero sequence voltage as transient characteristics after the validity analysis with the normal state and the applicability analysis of different fault conditions.Based on the comparative analysis of simulation examples,the validity and applicability of the proposed features are verified.Aiming at the problem of unbalanced data in distribution network,the improved negative selection method is used to identify the fault type.Immune idea can make full use of the "advantage" of huge difference between a large number of normal data samples and a small number of fault samples.However,the NSA has the problem of black holes and binary classification problem,so it is considered to introduce a small number of fault samples and combine with artificial immune network to optimize and cluster the detector.The simulation results under unbalanced samples show that the optimized detector set can effectively reduce the black hole range,and the fault state classifier generated by clustering can effectively reduce the invalid coverage range and realize the effective identification of fault types.Moreover,the method is still applicable to rising permeability of DG and different topologies.Aiming at the problem of small number of fault state data samples in distribution network,the improved DDQN method is used to identify the fault types.The problem of unbalanced samples is transformed into the problem of balanced small samples.Considering the reinforcement learning idea and the introduction of the sample dynamic label,the static fault type identification problem is transformed into a dynamic sequential decision problem.A second target network is introduced to approach the real label,so as to reduce the dependence on the label samples,and the double-layer network has different data enhancement effects on the original samples.The reinforcement learning identification network is designed to approach the action value function,which is used to select the action,update the state,and realize the sequential classification of samples;the termination judgment self reinforcement network is designed to approach the state value function,which is used to judge the termination state independently,so as to identify the same sample by multiple rounds.The simulation results under balanced small samples show that the self-reinforcement identification process can effectively reduce the probability of misjudgment.This algorithm can get more "humanized" output results through personalized customization,and the method is still suitable for rising permeability of DG and different topologies.Finally,the improved negative selection algorithm,the improved DDQN algorithm and the classical BP neural network algorithm are compared in algorithm idea,problem essence,mathematical model,data base and other aspects.