Stacking Ensemble Learning-Based Load Identification Considering Feature Fusion by Cyber-Physical Approach

Abstract

Nonintrusive load monitoring (NILM) methods considering multiple representative features can improve load identification. However, these existing methods assemble all features into one classifier, which easily causes overfitting and increases the computational complexity during the training process. To tackle these problems, this article proposes a novel stacking ensemble learning (SEL)-based load identification framework considering physical and cyber feature descriptors (CFDs), which can nicely fuse features by integrating different types of features into corresponding classifiers of the SEL model. The proposed framework includes two main parts: 1) comprehensive features are constructed based on a physical feature descriptor (PFD) and a CFD to fully explore the representative feature space and 2) the SEL model is developed to enhance the mutual complementary ability of different features and increase the variability of base classifiers. To evaluate the effectiveness of the proposed method, experiments are conducted on a public dataset. Numerical evaluations for residential building loads show that the proposed method significantly improves the identification performance and outperforms prior methods.