Multi-scale attention mechanism network for Nonintrusive load decomposition

Abstract

Abstract Nonintrusive load decomposition is an important prerequisite to realize intelligent power monitoring and a key application of the smart grid. The existing algorithms cannot achieve the decomposition effect with high accuracy and have poor performance in low-frequency loads. To solve these problems, a method based on a multi-scale attention mechanism network is proposed. First, construct the hole residual attention module to extract the deep features and help the network learn the important features of high peak areas and low-frequency electrical appliances in the time series data. Then, the multi-scale fusion module is proposed to fuse the characteristic information after the convolution operation of different scales. Finally, the decomposed active power values of multiple target electrical appliances are output through the full connection layer. The proposed work can decompose the load characteristics of each electrical appliance in the house by the nonintrusive method. The experimental results on the dataset indicate that this work can reduce the decomposition error and has excellent load decomposition capability compared with the existing methods.