Most of the existing non-intrusive load identification technologies are based on optimization and pattern recognition algorithms. Both algorithms have certain defects in model generalization ability and unknown load identification. In response to this problem, this paper proposes a non-intrusive load identification model based on metric learning. The load current characteristics are mapped into the metric space through a convolutional neural network, and the triplet loss is used during network training for feature aggregation. Then the load identification is realized by similarity discrimination of metric space features, which can realize effective recognition of unknown loads, and has strong generalization ability. On the other hand, metric learning is one of the methods of few-shot learning, which can reduce the dependence on training samples and has high practicability.