Feature Selection Method for Nonintrusive Load Monitoring With Balanced Redundancy and Relevancy

Abstract

Nonintrusive load monitoring (NILM) has become a key technology in the power Internet of Things as well as an important information source for load characteristics analysis. Whether the selected features are appropriate determines the effectiveness of the load identification in NILM. In order to reduce the redundancy and improve the relevancy of the selected features, a feature selection method that balances redundancy and relevancy is proposed. Based on the approximate Markov blanket decision, this article puts forward the basis of feature set redundancy elimination. This is used to determine the priority of feature selection by taking the number of features preselection as a measure. The mutual information method is combined with the CRITIC weight to implement redundancy elimination for the initial feature set of the load, and the feature correlation ranking algorithm is established based on the Relief-F method. Based on the idea of balancing redundancy and relevancy, a feature selection strategy is established for the initial feature set, in order to obtain the optimal feature subset with minimum redundancy and maximum relevancy. Finally, the <i>K</i>-nearest neighbor identification algorithm after <i>k</i>-value optimization is used to simulate the proposed method. The results show that compared with the other feature selection methods, the proposed method is promising in terms of robustness and generalization and shows an active effect on improving identification accuracy.