Abstract: Aiming at the problem that the current non-intrusive load identification algorithm fails to take into account the accuracy of load identification and the feasibility of deploying on embedded devices, a 1DCNN-BP load identification algorithm based on the idea of decision tree is proposed. First, in order to realize the extraction of load features and the dimensionality reduction of data features in the case of load combination switching, a two-stage event detection algorithm that can eliminate background load interference is designed, and a U–I space sequence feature extraction method based on curve description is proposed. Secondly, in order to have the generalization ability, high recognition rate, and the feasibility and economy of deploying on embedded devices, a 1DCNN-BP load identification method based on the idea of decision tree is proposed, which takes sequence features, load power and harmonic features as inputs. Finally, based on the Plaid and Blued-A public data sets, an example analysis is carried out. Under the condition that the required RAM and ROM are only tens of KB, the recognition accuracy rate reaches 92.3% and 100%, respectively, laying a solid foundation for the subsequent user-side energy management.