Nonintrusive Power Load Decomposition Based on Adaptive Graph Convolutional Neural Network

To fully exploit the correlation between the operating states of appliances, an adaptive graph convolutional neural network (AChebNet) for nonintrusive power load decomposition is proposed. An adaptive adjacency matrix is defined to characterize feature dependencies and uncover the hidden internal connectivity between features at different nodes in the graph model. This paper introduces the adaptive neighbor matrix to the Chebyshev Spectral CNN (ChebNet). By integrating a predefined neighbor matrix generated based on time intervals, we construct adaptive graph convolutions to better learn the graph structure and extract deeper hidden features. We explore the input dimensions of the model and select multiple relevant features based on the Spearman correlation coefficient to evaluate their impact on model performance. The proposed model outperformed ChebNet in experiments, achieving a 48.87% reduction in the mean absolute error (MAE) for the disaggregation of five appliances, and the mean power disaggregation accuracy improved from 87.39% to 92.74%. With multi-feature inputs, the model surpassed single-feature inputs, reducing the MAE by an additional 16.86% and increasing accuracy from 92.74% to 94.58%. Therefore, AChebNet can be effectively applied to reduce decomposition error and enhance overall accuracy in nonintrusive load decomposition.