| With increasing use of phasor measurement units in power system, an enormous amount of data is being stored in phasor data concentrators (PDCs). PDCs have the capability to store disturbance files separately. Over a period of time, the number of such disturbance files keeps increasing. However, these files are not really mined for data and mapped to actual events that may have caused the disturbance.;Identifying disturbance events recorded by phasor measurement units (PMUs) has drawn attention of researchers in recent times. Published literature documents some approaches to identify typical disturbance events frequently occurring in power system. However, in order to comprehensively identify all disturbance events recorded by PMUs, it is required to know how many types of events are detected and recorded by PMUs monitoring a certain power system. In other words, for classification purpose, one must know the number of classes to begin with.;This study uses the disturbance files stored from 2007 to 2010 inside the PDC owned by Public Service Company of New Mexico (PNM) to first cluster the files, mapping each cluster with a disturbance event. Knowing the classes (targets), it then examines the performance of four well known classifiers to classify the disturbance files into different types of disturbance events.;In order to perform the clustering, the data from disturbance files are first preprocessed. Features are extracted from these files using Minimum Volume Enclosing Ellipsoid (MVEE) method. After qualitative consideration of different clustering techniques, AGglomerative NESting (AGNES), a suitable unsupervised learning technique is chosen and successfully implemented. Results reveal that this process should be underpinned to any comprehensive event detection tool for PMU data.;Using the identified events after applying AGNES as targets, four most relevant supervised classification methods are implemented and tested to develop a disturbance identifier. These methods are Support Vector Machines (SVM), k-Nearest Neighbor Classification, Naive Bayes Classifier, and Recursive Partitioning and Regression Trees (RPART). Performance of these methods is quantified in terms of accuracy and speed. |
