Power System Fault Analysis Using Convolutional Neural Networks

With the progress of society,electricity has become an indispensable part of our life.Higher requirements for power system safety and reliability are ongoing challenges.In a power system,a small fault can cause a power system blackout that affects large areas,causing inestimable losses to the economy.Reliable analysis of power system faults is therefore an important basis for effectively protecting a power system.For this purpose a power grid must be capable of dealing with system faults,such as faults caused by equipment failure or wrong decision making,as well as faults caused by external influences such as natural disasters,a struck of lightning or a flood.As power system data becomes more and more available,it would be a waste of resources to not use the information stored in them to improve the quality of our power grid.Machine Learning techniques have enjoyed a wave of popularity during recent years and we nowadays have powerful algorithms at our disposal.They enable us to train neural networks to extract features from power system data in real time for decision making posterior to a power system fault.In this work,such techniques were used to analyze how convolutional neural networks perform for feature extraction on simulated power system data of different system architectures.The results achieved help to understand better how power system fault analysis can be approached using deep learning.Simulation studies have been made training networks with data from several generated fault scenarios(fault inception times,fault types,fault location)as well as data from systems with different loads and generation scenarios.Samples have been generated using Matlab and Simulink with Wavelet decomposition to extract meaningful frequencies for fault analysis.A sliding window approach is used to generate samples labeled ”faulted”or ”non-faulted”in order to use the data for supervised learning.