Non-Intrusive Load Decomposition Based on Graph Convolutional Network

Abstract

Demand-side power management and energy efficiency analysis are crucial to reducing energy consumption and improving power efficiency. Non-intrusive load decomposition is one of the important links to improve demand-side power management and energy efficiency analysis. In view of the fact that most of the current non-intrusive load decomposition methods focus on the analysis of traditional load characteristics and the optimization of algorithms, and lack of sufficient mining of user’s electricity behavior habits, a nonintrusive load decomposition model based on graph convolutional network (GCN) is proposed. The model firstly constructs the power sequence into graph data as network input based on the spectral graph theory, using the time characteristics extracted from the user’s electricity behavior habits. Then, based on the graph convolutional neural network, the power attribute features of each electrical appliance and its time-related structural features are extracted to achieve non-intrusive load decomposition. Specifically, it has a total of five layers of structure, including four layers of graph convolution layer and one layer of graph pooling layer. The ReLu activation function is used to improve the nonlinear expression, the dropout and L2 regularization measures are used to alleviate overfitting, and the bath size is used to improve the training speed. AMPds2 dataset is used for experimental testing. The experimental results show that the proposed decomposition model can accurately detect the switches of electrical appliances, and achieve better decomposition and effective tracking of the power of each electrical appliance.