Abstract: Non-intrusive load decomposition can be used to decompose the power of individual appliances from the total power reading of a household electricity meter, which has important practical significance for energy conservation and emission reduction, intelligent power consumption and so on. In view of the current problems of deep learning in load decomposition, such as low decomposition accuracy and large decomposition error of electrical appliances that are used infrequently, this paper proposed a non-invasive load decomposition model of sequence-to-point based on attention mechanism, which combines time convolutional neural network (TCN) and long short-term memory network (LSTM). Firstly, the power time series were fed as the network input in overlapping sliding window mode, TCN expansion causal convolution was employed to expand the receptive field of convolution kernel, the residual link and batch standardization were added to accelerate the efficiency of extracting deep load features. Meanwhile, the attention mechanism was introduced to extract key information, and then LSTM was employed to decompose the load decomposition by capturing the evolution mode of power series. Finally, the experimental results on UK-dale and REDD data set show that the model performs well and significantly improves the decomposition accuracy.