Abstract: Non-intrusive load monitoring can extract the operating state of a single load from the main data of a building, which is of great significance for consumers to adjust their electricity strategies and participate in demand-side response. Existing methods are limited by the sequential input of data in the Euclidean space, and cannot accurately describe the time correlation between different operating states of electrical appliances, which leads to unsatisfactory disaggregation accuracy. Therefore, a novel method of graph data modeling and graph representation learning was proposed. Firstly, the signal to be disaggregated was converted into power graph containing nodes and edges based on graph theory. Secondly, a graph convolutional network with residual mechanism was designed to fully mine the attribute features and time correlation features contained in the data at low sampling frequency. Then, a graph representation learning for load disaggregation was developed. Furthermore, an improved post-processing technology was proposed to solve the problem of the lack of refined correction strategies for dis-aggregation results, thus improving the overall performance of the model. Finally, the proposed method was validated by using public datasets AMPds2 and REDD. The results show that the method has a lower disaggregation error and better generalization performance.