Hybrid feature extraction and machine learning based optimized classifier for power quality disturbances

Abstract

In the past decade, the world has seen radical changes in the evolution of distributed energy sources and electric utilities. This has led to the proliferation of power electronic devices and nonlinear loads. The onus has been towards ensuring quality electric power to the electric utilities and its customer. This necessitates monitoring and analysing the electrical signals to understand the behaviour of the power network and initiate necessary mitigation action in the event of abnormalities to ensure the quality of power. This has generated interest in power quality research and so, this thesis focuses on efficient Power Quality (PQ) disturbance classification methods. The analysis of machine learning classifiers namely Extreme Learning Machine (ELM) and Support Vector Machine (SVM) towards classification of PQ disturbances is carried out. The features extracted from the S transform form the input to the classifiers. The results highlight the need to explore new approaches to improve the classification accuracy. In this perspective, this thesis suggests hybrid feature extraction technique combined with evolutionary computation based ELM classifier to enhance the classification in identifying the PQ disturbances. The hybrid feature extraction stage combines the capabilities of S transform and Hilbert transform to provide an efficient time frequency resolution. The feature vector is obtained by combining the features extracted from S transform and Hilbert transform. The feature vectors form an input to the machine learning classifiers. The analysis of machine learning classifiers highlights the capability of ELM classifier for accurate classification of the events. The major factor that decides the classification accuracy of the ELM includes the choice of input weight and hidden biases. This thesis proposes a craziness Particle Swarm Optimization (PSO) technique to optimally select the input weight and biases of ELM to ensure exactclassification. The results reveal the efficacy of the proposed algorithm in correct classification of disturbances. newline