Abstract: Load identification technology can quickly identify the types of electrical appliances,and plays an important role in household energy management,hazardous warning,response potential assessment and other aspects.Aiming at the problem that the existing load identification methods focus on the long-term or short-term single-scale load features,resulting in insufficient feature representation capabilities,and thus limiting model identification accuracy and generalization performance,a load identification and its interpretable interactive enhancement method based on multi-scale feature fusion is proposed.Firstly,shortterm features of high-frequency scale and long-term features of medium-and low-frequency scales are extracted from the load sampling data,and a deep feature extraction network with a double-tower structure is constructed to efficiently mine deep features of each scale by using different branches of the network.Secondly,a feature fusion network combining self-attention and crossattention is designed to fuse long-term and short-term features,improving the feature utilization of the model.Then,a training method of metric learning is used to shorten the feature distance of the same type of samples and improve the efficiency and effect of feature fusion.Finally,the interpretable analysis method based on gradient is used to quantify the importance of features,and realize the adaptive feature enhancement and model optimization combined with expert interaction.The experimental results show that the recognition accuracy and generalization ability of the proposed model are superior to those of existing models,and the interpretable analysis verifies that its effectiveness comes from the full use of multi-scale features.