Accurately identifying the type of voltage sag source is an important basis for both responsibility distinction between suppliers and users as well as governance. In view of the existing problems of difficult identification on type of voltage sag sources and low accuracy

a type identification method of voltage sag source based on the fusion of BiLSTM and CNN is proposed: Firstly

the voltage sag event data is subjected to discermment weighting and synthesis before being decomposed and reconstructed using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to filter out the effect of noise. Then

BiLSTM is used to decompose the cyclic features of the voltage sag time series signal to possible explore the temporal correlation between time series samples

extract global features

and generate feature vectors of unified length as much as possible. After that

CNN is used to further process and refine the spatial correlation of the feature vectors

emphasizing the local key features. Finally

the features obtained from both methods are fused to classify the sag source and determine the type of the sag source. The IEEE 33-node test system simulation model is set up to simulate voltage sag time and

on the basis of obtaining a large amount of simulation data including start - up of motor

no - load operation of transformer and various short - circuit faults

the proposed method is used for identification and differentiation on the type of voltage sag source and for comparison with the result identified respectively by the CNN and LSTM. The results show that the proposed method has high accuracy and good robustness.