Aiming at the problem that the current non-intrusive load disaggregation model based on deep learning is difficult to deploy at the edge with limited computing resources, this paper proposes a lightweight non-intrusive load disaggregation model based on the encoder-decoder structure. The model improves the accuracy of load disaggregation by introducing attention mechanism to calculate spatial attention and the improved channel attention separately. In addition, the design of different decoders is investigated in this paper, and the depthwise separable convolution is used to improve the residual block in the upsampling layer and reduce the number of kernels in the convolution layer, so that the model has fewer parameters and requires less computation while ensuring good load disaggregation performance. Tests are conducted on the public UK-DALE dataset to verify the load disaggregation performance of the proposed model and the feasibility of edge deployment.