Abstract: Non-intrusive load monitoring is one of the important technologies in the customer-side smart IoT scenario. In order to improve the accuracy of non-intrusive load decomposition and identification, this paper proposes a residential customer load decomposition method combining multi-scale attention mechanism and convolutional neural network. Firstly, the attention scores of the normal load data at the previous few moments of the attention model are smoothed dynamically against the abnormal scores at the current moment. The load identification attention model is optimized by constraint factors. Then, on this basis, convolution filters of different sizes are used to model the mixed load data of different electrical equipment, to mine more abundant characteristic information. Finally, taking the PLAID data set as an example, this paper compares the proposed method with the more popular load decomposition method. The experimental results show that the method based on the multi-scale attention mechanism in this paper can greatly improve the effect of load decomposition and reduce the confusion problem of electrical appliance identification with similar load characteristics.