Abstract: To reduce the waste of electricity in residents' lives, non-intrusive load monitoring shows its importance. Under the premise of event detection, a non-invasive residential load monitoring method based on convolutional neural networks-Bi-directional long short-term memory (CNN-BiLSTM) and dynamic time warping (DTW) is proposed. Firstly, the information of load operation state is measured by probability mass function, and the U-I characteristic curve of steady-state operation is extracted; Then, the image is normalized to a unified gray scale image, and the feature vector is extracted by convolution neural network as the load signature; Input the data into the BiLSTM for identification and use the DTW to optimize the identification results to achieve high identification accuracy. Finally, the PLAID public data set is used to simulate and verify the proposed algorithm. The simulation results show that the selected load signature has good identification performance, and the identification algorithm has higher reliability and accuracy than the comparison algorithm.