Abstract: Aiming at the fact that the accuracy of the traditional deep neural network disaggregation model still cannot meet the actual needs of non-invasive load monitoring, this paper proposes a load disaggregation model based on Temporal Convolutional Attention-based Network (TCAN). The model adopts the sequence-to-point disaggregation method, uses the improved temporal convolutional network as the basis to extract load data characteristics, increases the convolutional kernel sensing field, and obtains more data feature information. The model combines the attention module to extract richer and more valuable feature information, which improves the training efficiency. The experimental results in the UK-dale dataset show that the model has significant improvement in decomposing performance and judging the start-stop state of electrical appliances than the existing disaggregation methods.