Abstract: As the application scenarios of non-intrusive load monitoring technology continue to expand and load types become increasingly diversified, static identification methods based on single-layer features need to construct more comprehensive and complex features, and it is difficult to give consideration to both the accuracy and speed of load identification. A dynamic identification method based on multi-layer feature groups is proposed, which takes into account the sampling frequency, monitoring window width, computational complexity and load characteristic storage of different load features to construct a hierarchical feature group, extracts different feature groups for different load types in the load group as classification features to reduce the comprehensive extraction cost of features, and finally realizes on-demand load type recognition based on support vector machine multi-classification algorithm. Results from BLUED database simulation and actual engineering test of a factory show that the dynamic identification method based on multi-layer feature groups can increase the identification speed of most loads, while accurately distinguish among similar loads. It can give consideration to the accuracy and speed of load identification under the scenario of similar load and frequent switching, and has good engineering applicability.