The proposal of Electrical Internet of Things (EloT) has promoted the rapid development of smart electricity and load monitoring. In order to solve the problems that traditional non-intrusive load monitoring and decomposition (NILMD) is very time-consuming and has low identification accuracy, the semi-supervised learning (SSL) was used to cluster data and then establish characteristic data set, combined with load decomposition of fruit fly optimization algorithm with generalized regression neural network (FOA-GRNN). First, with the input of active power and current data, the similarity matrix was optimized with the method of semi-supervised learning. The operating state and power information of the electrical equipment were mined based on affinity propagation clustering. Then the operating state was represented as classification label with digital coding. Second, the time series data of the total active power, reactive power and current as well as the classification label matrix of the corresponding series were all input. To complete the model optimization and training, the optimal Spread value of the generalized regression neural network (GRNN) was obtained by using the optimization ability of the fruit fly optimization algorithm (FOA). Subsequently, the test time series data was input to obtain the classification matrix, that is, the operating status of each device, whose corresponding power information was utilized to reconstruct and fit the total active power in a way to complete the load decomposition. Through simulation comparison, the identification accuracy of the operating status of all electrical equipment is about 86% and that of single electrical equipment is about 96%. The proposed method is not time-consuming and significantly improves the mining ability and decomposition identification ability of load characteristics information.
