Power load disaggregation is to allocate the total power to each component electrical appliance reasonably. It plays an important role for users to understand their own power consumption habits, reduce power consumption and save electricity costs. To improve the accuracy and practicability of disaggregation, the main contributions are: (1) imported reactive power to establish multi-feature disaggregation model (2) Added penalty terms about the sparsity of component electrical appliance and working state conversion in the optimization objective function of the disaggregation model (3) proposed the effective working state identification results modification method. Compared with the new method on the public data set, the above method can effectively improve the accuracy of power load disaggregation. In addition, to reduce the required labor in the load modeling stage, unsupervised learning techniques were adopted to construct state power templates of features under different steady state conditions.