Non-intrusive load monitoring (NILM) refers to the method of installing monitoring equipment at the power entrance and using the total electrical load to obtain the status of individual electrical equipment on the power consumption side. This can accurately portray the power consumption portrait of users, so NILM is one of the key technologies for smart power distribution and fine management on the user side. With the application of deep learning in NILM, the ability of load identification and power decomposition has been improved, but the rate of training the model and the prediction accuracy of the model are still not high. For this reason, this paper proposes a load decomposition model based on the attention mechanism combining global and sliding windows. The model firstly maps the input total load power sequence to a high-dimensional vector through a power embedding matrix, and uses a BI-LSTM based encoder for information extraction; then, it selects from the extracted information by introducing an Attention mechanism combining global and sliding windows information with high correlation to the current moment, which is used for decoding, and finally, the power decomposition result is obtained. The data set REFIT is used to verify that the proposed algorithm has a better effect on speed and accuracy.