Recursive clustering K-nearest neighbors algorithm and the application in the classification of power quality disturbances

When identifying an unknown sample, k nearest neighbors (KNN) algorithm requires calculating the distances between this sample and all samples in the training set. It cannot satisfy high real time situation, such power disturbance recognition. We propose a novel method to reduce the size of the training set for KNN algorithm. The proposed method is based on iterative clustering. Clustering the original training set, if the cluster only contains one class, the samples in this cluster is represented by the cluster's center; otherwise, we cluster again. Finally, the original training set is represented by the subset which consists of all cluster's centers. Experimental results show that the classifier become much simpler after sample reduction. For the problem of power disturbance, only 6.28 percent of the original training set requires to be reserved while the accuracy has no significant difference.