Low frequency residential load monitoring via feature fusion and deep learning

Abstract

Non-intrusive load monitoring (NILM) is a technique used to disaggregate the total power signal into individual appliance power signals, which plays an important role in smart grid. Recently, deep learning is widely used to deal with the NILM problem. However, current deep learning models are purely data-driven, which do not consider physical mechanisms, making them less effective in extracting useful features. To address these issues, a new approach for feature extraction based on variational mode decomposition (VMD) and a new deep learning model based on variational autoencoder (VAE) are developed in this paper. The proposed feature extraction approach extracts the pulse feature and concatenates it with the original power data to form multiple features, i.e., which achieves feature fusion to improve the performance of deep learning models better than with a single feature. In addition, a feedback variational mode decomposition (FVMD) is proposed to improve the decomposition performance of the original VMD. The channel attention mechanism is introduced to VAE to improve the performance of the model. To verify the accuracy and robustness of the proposed scheme in NILM, it is compared with the state-of-the-art models on the UK-DALE dataset, and the results show that the proposed feature extraction approach can greatly improve the performance of deep learning models and the proposed new deep learning model outperforms some state-of-the-art models in the realm of NILM.