Abstract: Non-intrusive load disaggregation can accurately portray the user’s power consumption portrait by recovering the information of single electrical equipment on the power consumption side from the total electric meter information,which plays an important role in the refined management of the consumers. Aiming at the problems of low decomposition accuracy and low training efficiency of current artificial neural network models in load decomposition,this paper studies and builds a non-intrusive load disaggregation model based on temporal convolutional neural network. By analyzing the power consumption of the device,the dilated causal convolution is applied to perform convolution operations on the power sequence of electric meter and to expand the receptive field and extract richer features. The network training efficiency is improved by adding residual connections,weight normalization layers and optimizing training data window. Finally,the constructed model is tested on the optimized UKdale data set. The experimental results show that mean absolute error,root mean square error,and relative error are all in a relatively small range,and the time complexity analysis further shows that the model has a shorter training time without losing the load decomposition accuracy.