Abstract: To address the issues of insufficient single feature representation, missing power characteristics, and imbalanced datasets in non-intrusive load identification technology, this paper proposes a non-intrusive load identification method based on CNN-SENet and feature fusion. Firstly, the multidimensional feature information of the load is fused to enhance the feature expression capability by constructing a two-dimensional V-I trajectory map, a reactive current recurrence map, and an instantaneous active power grayscale map, which are mapped to the R, G, and B channels of the composite feature map respectively. Secondly, the ADASYN algorithm is used to balance the dataset and enhance the model's capability of identifying minority class samples. Finally, the SENet attention mechanism is introduced to combine with the convolutional neural network to improve the recognition capability of key features by optimizing the feature weight allocation, and the deep convolutional architecture is used to fully capture the local and global features of the image. The experiments are validated based on the PLAID public dataset, and the results show that the load recognition accuracy of the proposed method is 97.92%, and the F1 score is 97.91%, which is significantly better than the traditional method.