Efficient identification of the voltage sag sources is significant in the power quality studies. This paper presents a novel method for voltage sag source identification which performs automatic feature extraction and shows a superior performance regardless of the insufficient amount of training samples. In the proposed strategy, the input data are preprocessed and fetched into the feature extractor, which is designed based on the convolutional neural network. Then the weighted k-nearest neighbor classifier generates the identification results. In the training period, the few-shot learning technique is harnessed, and the siamese network is constructed such that the proposed model learns efficiently even with a small number of samples. The proposed scheme is implemented in Python and PyTorch framework. Case studies and comparisons with other methods are carried out on 700 samples of voltage sag events in Jiangsu Province, China. Experimental results show the superiority of the proposed method over other identification methods in the tested cases.