Abstract: Non-intrusive load monitoring (NILM) is a common method to research load information of resident users. However, it has some problems, such as low disaggregation accuracy and lacks of generalization and etc. Therefore, an NILM-alternating optimization (NILM-AO) method based on graph signal processing (GSP) theory was proposed. A graph signal model was constructed based on the total load data, and a power loss constraint can be obtained by the graph signal model to solve the lack of load data correlation research in traditional methods. Compared with the traditional method that requires altering model parameters, NILM-AO finds an optimal model parameter automatically, which improves the capability of real-time monitoring and decreases the operation cost of power grid. Simulation results show that NILM-AO improves the accuracy for 15%, and decreases the calculation time for 10% under 1-min sampling rate, which indicates the superiority of NILM-AO.