Development of an Artificial Neural Network based Hardware Prototype for Fault Localization in Distribution Grids

Abstract

In modern distribution grids, the task of fault localization using conventional techniques is increasingly becoming a challenge due to the rising domination of inverterbased, volatile, distributed power generation. Since approved methods from high voltage level such as reactance-based methods lack accuracy in distribution level topology, alternative approaches for accurate fault localization are required. Within the scope of this work, an artificial neural network (ANN) based solution for the localization of electric faults at distribution level has been developed, evaluated and implemented on standard hardware from industrial automation technology i.e. a programmable logic controller (PLC). A reduced yet representative model of a distribution grid incorporating a variety of aspects influencing the accuracy of fault localization such as distributed generation, ring network topology with open or closed loop as well as variable fault resistance has been developed. Current and voltage measurements generated under various fault conditions have been used for training of an ANN. Different ANNs have been trained with various network structures and training algorithms and after thorough analysis and comparison of their performance, the most suitable networks have been implemented on hardware and tested in hardware-in-the- loop configuration. Thereby a real-time simulator suitable for application testing and rapid prototyping provided process values of the modeled distribution grid.