Non-intrusive load identification decomposes the power usage of each appliance to achieve precise monitoring of appliance behavior by analyzing the total power consumption data of a household or enterprise. It is of great significance for energy conservation, reducing power costs, and promoting smart grids. When the current load identification methods deal with complex time series data, it is difficult to capture the scale diversity of the long-term and short-term dependencies of the data, resulting in limited identification accuracy. In response to the above problems, the multi-scale features of time series data are studied. Through a multi-scale feature extraction strategy, the information at different time scales in the data is effectively captured. Meanwhile, the potential attention mechanism is introduced into the extracted multi-scale features, enabling the model to focus on the key moments and important features in the data while reducing the computational complexity. Finally, experiments are conducted on the public dataset. The results show that, compared with other comparison methods, the proposed model achieves the best effect, verifying the validity of the model.