Combining machine learning techniques with Kappa–Kendall indexes for robust hard-cluster assessment in substation pattern recognition

Abstract

This study proposes a method that combines different machine learning and lean six sigma techniques to calibrate cluster analysis through linkage methods. The power quality indexes of substations in Brazil, which are of interest to regulatory agencies, are used. The method uses the random forest mixed with rotated factor analysis to filter, minimize, and improve the interpretation of latent information. Variability scenarios are created using the Monte Carlo simulation to assess the stability of the cluster analysis using the design of experiments and the Kappa–Kendall indexes. The Ward method shows a better consistency in all scenarios and a better discriminatory power among the clusters. The optimal result is used to predict different scenarios with high levels of variability (5, 10, and 15%) by comparing the behaviors of different supervised machine learning techniques for classification. The results show that the k-nearest neighbors, support vector classifier, and logistic regression approaches can accurately predict, even in scenarios with high variability in the dataset.