Feature selection of Complex Power Quality Disturbances and Parameter Optimization of Random Forest

Abstract

A hybrid method based on modified discrete artificial bee colony algorithm(MDABC) for power quality disturbance(PQD) signal feature selection and parameter optimization of random forest(RF) is proposed. Firstly, time-frequency features of the complex power quality disturbance signal are extracted using s-transform(ST) to form an original feature set; Then, by using default parameters of RF, the out-of-bag(OOB) Permutation test value of each feature in the original feature set is calculated as the feature weight, and the features of the training and verification data set are rearranged in descending order accordingly. Finally, taking the generalization error of RF as the objective function, the forest scale: nTree, the number of input features:Ni and the node feature subset size:q are optimized using MDABC to determine the optimal parameters of RF and the optimal feature set. It can be seen from the experiment that compared with the RF classifier before optimization, the accuracy of the MDABC-RF classifier in the classification of 16 and 19 complex PQD signals is better, and its operating efficiency is greatly improved.