One-Class Classifier Based Fault Detection And Localization In Distribution Systems With Distributed Energy Resources

With the development of the smart grid and the continued expansion of the installed capacity of renewable energy,the issue of integration of renewable energy into power systems has attracted much attention.These energies highly dispersed in distribution systems are called Distributed Energy Resources(DER).And the large-scale integration of DERs into the power grid poses a daunting challenge to the reliability and stability of power systems,especially to the protection system.Therefore,it is really urgent to detect and localize the fault in distribution systems with DERs.Recently,the advancement of the power grid and IoT technologies have provided massive sensing and monitoring data,which makes it possible to apply data-driven methods to detect and localize system faults.Considering that the power grid runs at normal conditions for most of the time and the number of possible fault conditions increases exponentially with the size of the grid,it is difficult and impractical to collect the comprehensive set of sensing data under all possible fault conditions.Therefore,this thesis uses only the normal data to train a machine learning method,namely the one-class classifier,for the detection and localization of faults in distribution systems with high penetrations of DERs.The research contents of this dissertation are as follows:First of all,in view of the problems of relay miss-trips and mal-operations when DERs are integrated into distribution systems,we adopt the Support Vector Data Description(SVDD)algorithm to detect faults in power systems.Meanwhile,aiming at the problem of parameter selection of SVDD,a hybrid algorithm for narrowing the search range of parameters is proposed.Compared with the traditional relay protection scheme,the SVDD-based offline detection model can significantly reduce relay miss-trip and mal-operation rate caused by the high permeability of DERs.In addition,the proposed hybrid algorithm has lower training overhead with improved SVDD performance.Secondly,in view of the concept drift problem caused by fluctuating penetration level of DERs in power grid,this thesis uses the Incremental Support Vector Data Description(ISVDD)algorithm to online identify faults in distribution system.Aiming at the slow performance improvement of ISVDD online training,the ISVDD detection model combined with the Hyperspherical Data Augmentation algorithm(HISVDD)is proposed.Compared to ISVDD,HISVDD online fault detection model has significant performance and training speed improvements when the appropriate hypersphere data distribution radius and data distribution are selected.Finally,aiming at the problem of coarse resolution in terms of the region selected for fault detection using classification methods,this thesis proposes a fault localization model based on fault probability of sub-regions according to the partition criterion of distribution networks.The model combines the Kernel Density Estimation(KDE)to indicate the fault confidence level of each sub-regions by P value,and then accurately locate the fault source by comparing the P values,which can effectively cope with the inaccurate positioning problem under the classification method.